.1»: :1 .._
4.

.....%
u- ..r. ‘

.41

.1

Ah

W r 1.993

7 .4. tag GM“
. ma .. o . r
.r...&. #5qu
.5 . 3. ...Z
. .2 .

r:
”humuﬂuw
i)...
.332,”

”an;

$.13. 14

.o. .

FwMian

1.2»
a.

 

A . Hutuﬂnw. ..

{1 are] . , ‘ . Mwmﬁmmﬁ ‘n.
. .. u if}. , 1.1352 ‘ . ‘ ,
Q . -. V 90.13932: .51“ 3‘ x. ‘ T.
inﬁll}! 2......‘0’ I.\¢.\.ol.... .331. an»: ‘
ll.
,Il

 

THEOlS

.‘\ I
”J 1" o F
k”

22033/7/(005

This is to certify that the
dissertation entitled

AN ECONOMETRIC STUDY OF THE EFFECTS AND
MOTIVATION FOR CENTRAL BANK INTERVENTION IN

PhD.

TURKEY: 1993-2003

presented by

PINAR OZBAY

has been accepted towards fulﬁllment
of the requirements for the

degree in A ECONOMICS

2 r »-

W

Major Professor's Signature
ngurl 1&a004

Date

MSU is an Aﬁnnative Action/Equal Opportunity Insﬁtuﬁon

. - ,.— --"‘-‘—.'_'. " _

.__._

4 '-

 

LIBRARY
Michigan State
University

 

 

 

PLACE IN RETURN Box to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 cJCIRC/DateDuepes-sz

AN ECONOMETRIC STUDY OF THE EFFECTS AND MOTIVATION FOR
CENTRAL BANK INTERVENTION IN TURKEY: 1993-2003

by

Pinar Ozbay

A DISSERTATION

Submitted to
Michigan State University
In partial fulﬁllment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY
Department of Economics

2004

ABSTRACT
AN ECONOMETRIC STUDY OF THE EFFECTS AND MOTIVATION FOR
CENTRAL BANK INTERVENTION IN TURKEY: 1993-2003
by

Pinar Ozbay

This dissertation consists of four distinct chapters, all of which model and examine
the effects and motivation of central bank intervention in Turkey during 1993-2003.
Chapter one considers the effectiveness of central bank intervention in the Turkish Lira-
US$ spot exchange rate market for different exchange rate policy regimes in Turkey.
Forms of GARCH models are used to present the effect of intervention on both the mean
and volatility process.

Chapter two is concerned with the motivations of the central bank intervention in the
Turkish Lira-US$ spot exchange market. The central bank intervention reaction function
is modeled using probit, tobit and censored least absolute deviation models.

Chapter three examines the intervention reaction function with a different approach.
The Threshold model is used in the non-linear estimation of the reaction function in the
Turkish Lira-USS spot exchange market.

Chapter four is concerned with the relation between risk premium and central bank
intervention. Forward rates are calculated for the Turkish Lira-US$ spot exchange

market and then the effect of central bank intervention on risk premium is presented.

Cepyright by
PINAR OZBAY
2004

This dissertation is dedicated to my beloved husband, Gokce Ozlu, my mother, Hatice
Ercan, my brother and my sister
For their love, encouragement and trust

iv

ACKNOWLEDGMENTS

Without the assistance and encouragement of many people, I could not have ﬁnished
this work and present it in this form.

First of all, my sincerest thanks goes to Professor Richard T. Baillie, my mentor. His
technical and general advice made my dissertation possible. He gave me valuable
comments and suggestions . His help and comments were invaluable.

I owe a special thanks to Dr. Ana Maria Herrera. She gave me very helpful
comments for my dissertation and helped me to complete it.

I like to thank Dr. Thomas Jeitschko for his advice. He helped me to navigate the
rough waters of the whole dissertation process.

I also like to thank Artem Prokhorov and Jongbyung Jun for their advice and
friendship. Especially Artem Prokhorov, who gave me great help in studying
econometrics and the GAUSS programming language.

The greatest thanks, however, goes to my lovely family. My beloved husband, Gokce
Ozlu, has been a constant source of inspiration and encouragement. His dedication and
support have been so great and helped move my dissertation forward.

And my family have always supported and encouraged me in everything that I have
done. Their love, patience and trust are things that I would never be able to repay. Thanks
to all of them.

Finally, I gratefully acknowledge support from the Central Bank of Turkey and the
Department of Economics, Michigan State University. They provided me with the

scholarship and helped motivate me to complete my dissertation.

TABLE OF CONTENTS

LIST OF TABLES ................................................................................. viii

LIST OF FIGURES .................................................................................... x

CHAPTER 1. IS CENTRAL BANK INTERVENTION EFFECTIVE?

1 .1 . Introduction ............................................................................... 1
1.2. Economic Developments in Turkey ................................................... 11
1.3. Data ...................................................................................... 32
1.4. Model and Estimation Output ........................................................ 43
1.5. Conclusion .............................................................................. 54

CHAPTER 2. WHAT MOTIVATES CENTRAL BANK INTERVENTION?

2.1 . Introduction ............................................................................ 55
2.2. Modeling Intervention Behaviour ................................................... 57
2.3.1 Probit Model ......................................................................... 60
2.3.2 Estimation Results .................................................................... 63
2.4.1. Tobit and Censored Least Absolute Deviation Model ......................... 67
2.4.2. Estimation Output ................................................................... 69
2.5. Conclusion ............................................................................... 76

CHAPTER 3. THRESHOLD NON-LINEAR ESTIMATION OF THE CENTRAL

BANK INTERVENTION RACTION FUNCTION

3.1 . Introduction .............................................................................. 78
3.2. Estimation of Threshold Model ....................................................... 79
3.3 Tests for Nonlinearity and Estimation Output of Reaction Function ........... 86

vi

3.4. Conclusion ............................................................................. 98

CHAPTER 4. RISK PREMIUM AND CENTRAL BANK INTERVENTION

4. 1 . Introduction ............................................................................. 99
4.2. Details of the Model: Risk Premium and Intervention ......................... 101
4.3. Data .................................................................................... 104
4.4. Estimation Output .................................................................... 111
4.5. Conclusion ............................................................................ 114
CHAPTER 5. CONCLUSION ................................................................... 1 15
LIST OF REFERENCES ......................................................................... 118

vii

LIST OF TABLES

CHAPTER 1

Table 1.1: Main Economic Indicators ............................................................ 13
Table 1.2: Estimation of Intervention Model (1993-1999)......................................47
Table 1.3: Estimation of Intervention Model (1995-1999) .................................... 48
Table 1.4: Estimation of Intervention Model (2000-2003). ...................................... 52
Table 1.5: Estimation of Intervention Model (2001-2003). ........................................ 53
CHAPTER 2

Table 2.1: Estimation ofProbit Intervention Model (1995-1999). 64
Table 2.2: Estimation of Probit Intervention Model (2001-2003). . . . . . . . . . . . . . . . . . . . 6........6
Table 2.3: Estimation of Tobit Intervention Model (1995-1999) ............................. 71
Table 2.4: Estimation of Censored LAD Intervention Model (1995-1999). . . . . . . . ...........72
Table 2.5: Estimation ofTobit Intervention Model (2001-2003). . . . . . . . ...74
Table 2.6: Estimation of Censored LAD Intervention Model (2001-2003) ................. 75
CHAPTER 3

Table 3.1: Test for Non-Linearity (1995-1999) ................................................. 87

Table 3.2: Estimation of Threshold Intervention Reaction Function (1995-1999) ........... 89
Table 3.3: Least Squares Estimation of Conditional Variance (1995-1999). . . . . . . . . ....90
Table 3.4: Test for Non-Linearity (2001-2003) ................................................. 92
Table 3.5: Estimation of Threshold Intervention Reaction Function (2001-2003) ........... 95

Table 3.6: Least Squares Estimation of Conditional Variance (2001-2003). . . . . . . . . ...96

viii

CHAPTER 4
Table 4.1: Estimation of Intervention/ Risk Premium Model (1994-1999). . . . . . . . . .....1 12

Table 4.2: Estimation of Intervention/ Risk Premium Model (2001-2003). . . . . . . . . .....1 13

ix

LIST OF FIGURES

CHAPTER 1

Figure 1.1: Percentage Change in Annual GNP ................................................. 14
Figure 1.2: Inﬂation and Interest Rate .............................................................. 15
Figure 1.3: Spot Rate (TL/USD) (November 1993-1994) ....................................... 18
Figure 1.4: Overnight Interest Rates (November 1993-1994) .................................. 19
Figure 1.5: Spot Rate (TL/USD) (1995-1999) ................................................... 22
Figure 1.6: Spot Rate (TL/USD) (2000-2003) .................................................... 26
Figure 1.7: Overnight Interest Rates (J anuary-October 2000) ................................. 27
Figure 1.8: Overnight Interest Rates (March,2001-2003) ....................................... 28
Figure 1.9: Spot Return (TL/USD) (November 1993-2003) .................................. 34
Figure 1.10: Spot Return (TL/USD) (November 1993-1994) ................................. 35

Figure 1.11: Overnight Interest Rates (November 1993-1994)................................36

Figure 1.12: Spot Return (TL/USD) (1995-1999) ............................................... 37
Figure 1.13: Overnight Interest Rates (1995-1999) .............................................. 38
Figure 1.14: Spot Return (TL/USD) (2000-2003) ................................................ 40

Figure 1.15: Overnight Interest Rates (January, 2000-October 2000).... . ....41

Figure 1.16: Overnight Interest Rates (March, 2001-2003)........... . ...42

CHAPTER 4

Figure 4.1: Forward Exchange Rate (TL/USD) (1993-2002) ................................. 106
Figure 4.2: Spot Exchange Rate (TL/U SD) (1993-2002) ..................................... 107
Figure 4.3: Forward Rate Forecast Error (1994-1999) ........................................ 109
Figure 4.4: Forward Rate Forecast Error (2001-2003) ....................................... 110

CHAPTER 1

Is Central Bank Intervention Effective?

1.1 Introduction

Exchange rate intervention in the foreign exchange market is deﬁned as the buying
and selling of foreign currency by a central bank or government in an attempt to
inﬂuence the level of an exchange rate. There are two main motives of intervention. The
ﬁrst one is to inﬂuence the level of an exchange rate and the second one is to reduce its

volatility or in other words to calm a disorderly market.

Intervention is generally sterilized such that purchase (sale) of foreign currency is
offset by a corresponding sale(purchase) of domestic government debt to eliminate the
effect on the money supply. Hence intervention is quite distinct from pure monetary
policy. Sterilized intervention provides monetary authorities to aim an exchange rate

target objective independent of their monetary policy.

Studies of intervention often distinguish between sterilized intervention, which does
not affect the money supply, and non-sterilized intervention which does. Despite the
agreement among economists about the effectiveness of non-sterilized intervention, the

effect of sterilized intervention is much more controversial. Non-sterilized intervention is

expected to directly affect the exchange rate because it changes the stock of base money
and involves broader money aggregates, interest rates, real demands for goods and assets,
and market expectations. Sterilized intervention leaves the monetary base unchanged. So
it is not expected to affect the exchange rate through the monetary channel. In contrast to
the direct effect, intervention, even if sterilized, can inﬂuence exchange rates through the
portfolio and the signaling channels. Sterilized intervention will alter the relative supplies
of domestic money and bonds. With risk averse investors who view domestic and foreign
bonds a s imperfect s ubstitutes, the impact 0 f intervention will In ake them trade b onds
which, in turn, will lead to an adjustment of the relative rate of return by changing the
exchange rate. A sterilized purchase of foreign exchange increases the amount of publicly
held domestic bonds, relative to foreign bonds, and induces a depreciation of the

domestic currency.

The portfolio balance theory implies there will be no impact of intervention on the
exchange rate when there is perfect substitutability of bonds. To test this, theory requires
information on the relative supplies of the assets. In the signaling channel, central banks
try to inﬂuence exchange rates by conveying to the market a signal concerning current or
future monetary or exchange rate policy. That is, sterilized purchases of foreign currency
are expected to lead to a depreciation of the exchange rate if the foreign currency
purchase is assumed to signal a more expansionary domestic monetary policy. This
channel is largely acknowledged as the main path by which interventions affect foreign

exchange markets.

There is an enormous amount of literature concerning the effects of intervention on
currency markets and the motivations for intervention. While researchers have been able
to document short-lived effects, the overwhelming conclusion after studying the literature
is that foreign exchange intervention is not effective for inﬂuencing the level and the
volatility of the exchange rate. See Edison (1993), Almekinders (1995), Baillie et a1.
(2000), Schwartz (2000), Samo and Taylor (2001), King (2003) and Humpage (2003).
Empirical studies of the 19803 are characterized by two major handicaps, a lack of data
on intervention and also a lack of survey data on exchange rate expectations. Studies on

intervention in the 1990s and early 2000‘s use high quality daily data on intervention.

Policy-makers seem to believe that it is possible for a central bank to signiﬁcantly
affect the supply and demand conditions in the foreign exchange market. In the 19803
and 199OS, attention focused on the effect of sterilized intervention on the level of the
exchange rate and the channels through which intervention works. Humpage (1988) and
Obstfeld (1989) note that the monetary amount of intervention is very small in
comparison to the total market trading. Most empirical studies, such as Humpage (1988),
Baillie and Osterberg (1997b) and Beine et a1 (2002) ﬁnd a “leaning against the wind”
phenomenon, which is regarded as economically unsatisfactory. Usually if any signiﬁcant
relationship is found, it is of the “leaning against the wind” type, with the Fed buying
dollars associated with a subsequent dollar depreciation next period. This is almost
certainly due to the policy endogeneity issue, with the bank buying dollars to support the

already depreciating currency.

 

 

 

There is general consensus in the literature that intervention does not affect exchange
rates through the portfolio channel. Lewis (1988) estimates outside bond demand
equations from the portfolio balance model of exchange rate determination for ﬁve
currencies. The portfolio balance model focuses upon bonds ﬁom outside the economic
sector that arise from the government debt. Different from the previous studies, she
decomposes the foreign asset by currency. Despite improved empirical techniques,
estimates of the portfolio balance model remain imprecise. Ghosh (1992) tested the
portfolio balance channel for sterilized intervention by examining the effects of changes
in relative asset supplies on the dollar-deutschmark rate during the 19805. A weak but
signiﬁcant portfolio balance inﬂuence is found. He concludes that substantial
intervention is required to inﬂuence the exchange rate through portfolio balance effects.
Dominquez and Frankel (1993b) estimate a portfolio balance equation that is consistent
with mean-variance optimization. In this model, expected ﬁrture exchange rates are
represented by data ﬁ'om surveys of expectations among private exchange market
expectations. Their main conclusion is that sterilized interventions are effective if they
change the risk premium and that cumulative intervention has a statistically signiﬁcant

effect on the risk premium.

Edison (1993) provides a comprehensive review of literature that focuses on
empirical work conducted in the 19808. Most of the empirical evidence summarized in
Edison‘s survey suggests that intervention can affect the exchange rate through the
signaling channel but not the portfolio-balance channel. Almekinders (1995) agrees with

Edison (1993) that no systematic effect of sterilized intervention via the portfolio

balance channel is found in empirical evidence. Data limitations and theoretical and
econometric problems have made it impossible to estimate the portfolio balance model
and measure the effects of sterilized intervention satisfactorily. Only ofﬁcial exchange
market operations, which c reate e xpectations o f c hanges in m onetary p olicy, o r w hich
embody another sufﬁcient ‘news’ content, appear to have a chance of affecting the
exchange rate signiﬁcantly. In their brief survey on recent literature, Samo and Taylor
(2001) focus on the empirical work conducted in 19905. They suggest that the portfolio
balance channel will diminish in importance over time, at least among the major
industrialized countries, as capital markets become increasingly integrated and the degree

of substitutability between ﬁnancial assets increases.

Most monetary authorities and economists believe that if intervention works, then it
must be through the signalling or expectations channel. See Samo and Taylor (2001). The
signaling channel assumes that intervention affects exchange rates by providing the
market with new and relevant information to other market participants. In research on
signaling, coordinated intervention is assumed to be more. effective than unilateral
intervention. Despite the consensus among economists about the effectiveness of
coordinated intervention, there is still a debate on whether or not the transparency of
monetary authorities‘ actions and objectives is necessary. While some studies provide
that secrecy is desired, others recommend that intervention be made public and the
objectiveness made transparent in order to increase the effectiveness through the
signaling channel. In a 1990 study, Dominquez discusses credibility games involving

signaling and concludes that credible intervention signals, will be effective as the non-

sterilized intervention when the central bank invokes its promised monetary policy. The
study also implies that coordinated intervention may differ from unilateral intervention in
terms of its impact on foreign exchange markets. Coordinated intervention has a
signiﬁcantly different and longer-term inﬂuence on market expectations than does
unilateral intervention. Weekly data on monetary surprises, exchange rates and
intervention is used and the effectiveness of intervention to vary with the credibility of
monetary policy is found. Similar to Dominquez (1990), Humpage (1999) found that
coordinated intervention increases the probability of success in affecting market
expectations. The size of an intervention also affects its probability of success, but
coordinated is generally a better predictor. He tested the hypothesis that the Federal
Reserve routinely has better information concerning exchange rate movements than the
market following the Louvre Accord. The frequency of success was relatively low,
implying that the US interventions generally possessed little forecast value. However,
there was evidence of intervention as a predictor that recent exchange rate movements
would moderate, but not reverse. Coordinated and possibly the amount of intervention
increased the probability of success. Samo and Taylor (2001) mention that ofﬁcial
intervention can be effective through its role as a signal of policy intentions, especially
when it is publicly announced and coordinated. Contrary to Dominquez (1990) and
Humpage (1999), Vitale (1999) shows that when the uncertainty about the objective of
foreign exchange intervention is particularly severe, the monetary authorities may target
the exchange rate more successfully. The market is more efﬁcient when this objective is

secret than when it is common knowledge.

Baillie, Humpage and Osterberg (2000) surveyed the literature on informational
issues. They suggest researchers should carefully consider whether intervention might
provide information to the market or whether the authorities have a clearer underaanding
of market conditions than less informed private traders. Popper and Montgomery (2001)
discuss that a central bank can affect the exchange rate by aggregating and disseminating

agents‘ information. The intervention is a useful way to transmit this information.

Most of the empirical studies show that sterilized interventions have little impact on
the exchange rate. Fatum and Hutchison (2003) used an event-study approach to assess
the effectiveness of intervention operations. In contrast to other studies, they found that

ofﬁcial intervention is effective when directed to short-run objectives.

An often stated objective of intervention policy is to cahn disorderly markets. Bonser-
Neal and Tanner (1996), Baillie and Osterberg (1997b), Hung (1997), Chang and Taylor
(1998), Dominquez (1998), Beine et a1 (2002) all ﬁnd evidence that intervention tends to
increase spot exchange rate volatility.Dominquez (1998) mentions that while intervention
policy oﬁen inﬂuences exchange rate volatility, it is not volatility that causes
intervention. A more surprising result is that secret interventions were generally found to
increase volatility. This result provides evidence that the more ambiguous are signals, the
more likely they are to increase volatility. Galati and Melick (1999) focused on
intervention that is perceived by market participants, different from the actual
interventions. Their analysis suggests that, on average, perceived intervention in support

of the dollar fails to strengthen the currency. They ﬁnd that market perceived intervention

may increase the uncertainty prevailing in the market regarding future movements in the
spot market. Besides these two effect, perceived intervention does not change the
balance of weights that market participants assign to future sharp appreciations and sharp

depreciations.

There is empirical evidence which indicates that some types of intervention can affect
the risk premia in forward markets. However, the risk premium is not necessarily the
intended target of the intervention .Likewise, evidence that spot exchange rate volatility
is frequently increased following intervention may be an unintentional extemality of
intervention. Baillie and Osterberg (1997a) ﬁnd that purchases of US dollars by the
Federal Reserve appear to have signiﬁcantly increased dollar denominated returns over
uncovered interest rate parity for both the DM-$ and the Yen-S. However the effect is not
symmetric since the Fed‘s actions has an impact, while the other central bank‘s actions
do not. They also ﬁnd some evidence that intervention leads to increases in volatility of
the risk p remium, as w ell as a ffecting the c onditional m ean o f the risk premium. The
nature of the causality appears unidirectional since there is no evidence that volatility of
the forward premium Granger causes intervention. Consistent with this study, Baillie and
Osterberg (2000) ﬁnd some support for the intervention variables affecting the risk
premium in the analysis where the relationship between daily deviations from uncovered
interest rate parity and intervention is investigated by using daily overnight euro-currency

deposit rates.

Another area of investigation has been the relationship between central bank proﬁts
and intervention. Sweeney (1997) noted that if central banks have better information than
the market, then they should consistently be able to earn proﬁts on their intervention
activity. Leahy (1995) and Neely (1998), Saacke(2002) had similar results. Kim and
Sheen (2002) note that proﬁtability is important in determining intervention like the other
factors s uch a s, o vemight interest differentials, volatility smoothing and exchange rate

trend correction.

Following Samo and Taylor (2001), King (2003) and Humpage (2003) present a brief
survey on the empirical work conducted post-1992. Similar to Sarno and Taylor (2001),
Humpage (2003) mentions that if the transaction of the given intervention is large and
coordinated, the probability to have the desired effect increases. He emphasizes the
methodological problems of the estimation techniques and suggests that the timing issue
with respect to the intervention data should be solved; otherwise empirical studies will
give limited information about the efﬁcacy of intervention. He suggests using high
frequency and intra-day data, so that intervention appears to affect the exchange rate
within minutes. Empirical studies using intra-day intervention data are available in the
recent literature. See Dominquez (2003) and Chang and Taylor (1998). Dominquez
(2003) used intra-day (5-minutes) intervention data on DM-USD and Yen-USD. She
mentions that intervention operations, especially coordinated types, are consistently
associated with increases in intra-day and daily volatility while there is little evidence
that interventions inﬂuence longer-term volatility. Chang and Taylor (1998) also ﬁnd

similar results. King (2003) mentioned that empirical studies show that foreign exchange

intervention has a transitory effect on the level or volatility of the exchange rate because
foreign exchange intervention may be undertaken to meet a range of objectives and the

studies do not adequately adress how these objectives may vary over time.

The results on the effectiveness of intervention are mixed and depend on which
exchange rate regime is analyzed, what sample period is studied and on the intervention
strategy that is followed. Empirical studies are often done for the developed and big
economies, yet few studies are available for developing and small economies. Hutchison
(2003) discusses that coordinated intervention is desirable in the short-run stabilization
policies of developing countries. Data on the amounts of intervention used by countries
with their currencies in target, or managed exchange rate regimes are usually impossible

to obtain and are regarded as highly secret, unlike the data on free-ﬂoating regimes.

In this study, the effect of intervention is investigated in Turkey. It is a small
economy that has had high inﬂation during the last 20 years. There is only one empirical
study in recent literature with respect to F X intervention in Turkey. See Domac and
Mendoza (2004). They ﬁnd that interventions decreased the volatility of exchange rate
during the free-ﬂoat regime covering the period February 22, 2001 through May 30,
2002. This study used the data which covers the period November 1993 through
December 2003. The Turkish economy had both managed and free ﬂoat exchange rate
regimes during this period. This enables us to make a comparison of the effect of
intervention under different exchange rate regimes in a small-open economy with high

inﬂation. Section 1.2 discusses the economic developments through the whole period

10

covered in this study. Sections 1.3 and 1.4 discuss the data, model and the estimation

results. Finally Section 1.5 gives a brief conclusion.

1.2 Economic Developments

Turkey adopted liberal economic policies in 19805 and ﬁnancial liberalization
brought important changes in the institutional arrangements of ﬁnancial markets . Most
trade liberalization was completed in the mid-19805 and Turkey entered into a customs
union with the European Union in 1996. Similarly, the reforms in the ﬁnancial area were
also rapid and decisive. Interest rates were liberalized in the early 19805, residents were
permitted to hold foreign exchange (FX) deposits in 1984. New institutional
arrangements helped to develop money and bond markets in the late 19805, and capital
movements and the exchange rate system were fully liberalized in August 1989. Despite
the measures taken in the ﬁnancial sector, public ﬁnance did not receive its proper share
of reforms in the 19805. The economic environment was very unstable throughout the
19905 contrary to the developments throughout the 1980s. The ﬁnancial crises, external
shocks, high inﬂation, lack of ﬁscal disipline, the difﬁculty in the rollover of the growing
domestic debt and political uncertainty were the main problems that created the unstable
environment. Stabilization programs targeting price stability did not have political

support, so efforts to decrease inﬂation were not successful.

11

The economic growth rate has been quite high except the crisis periods in 19905.
Both the nominal rates of interest and inﬂation were very high almost throughout the

decade of the 19905. See Table 1.1, Figure 1.1 and Figure 1.2.

12

 

4>m_.m in 2.9.: moose—an 52883 3° & 023

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

$3 33 3o“ 3: 3mm 38 $3 .38 $8 88 82 N8» MS“
02.": i i M: E E 2 3 3 .3 $_ -3 A 29
53:83 2: u: 8E 32 8N u: 8... 8a 8m 82 $0 as. MS
533353] m: 8.0 mi 32 E a; :3 a; “3 8a 8... m3.
n==oa>noocawma=8 No -2 .2 0.0 3 -.ﬁ 2
02:1 me; ”@823 as S . mm 3 n5 SJOT
3mm} 3 g $— 3 m: as Ew— aa 3.... aa—

 

 

 

 

 

 

 

 

 

 

 

moSoQ 08:2 5% 01:6 moccga a .3283 man 236 06358:. mama 59.83 a magmas

3 08mm zmaoag 38:2 623 03,5: mam
3 93% 5 253% o: :2: u as
1.5822 >58: >52QO acacia mma

...... 96:0 mmo§ acacia $358.83
a 3683
3 388m"

13

 

_ Hmmcwc HL" Hun-6939mm" Oran—mo m: >==:m_ 02W

 

7.\/

3 L
w
_
_

 

 

..7 3
. /\ / \

 

/ > \ 7,

 

 

 

.. . _ i <

<

 

 

rum

H

. /\/<\ L

 

 

 

 

 

 

waa ubWu HOWN wau wah Hwem uwwa HOW.“ wam waw MAE: Nag: NQGN Ncau

 

 

AJ 03% Zeno—5— 5.252 A023 m: 0.5.2: 5.38

14

L uni—=6 ab" main-ac: wan mane—den ”v.03 L
_ um: um: .

He: :! li||11-- ,- -,;|l-irli-aiilr,rl , -..Hac

 

 

 

Ha: r-.iiii:.|:llg gull»; ,z1:;r - ,:- ii: use
HNG 1 ,
we: r
L a:
V a: -

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

L

we . . _ i . . . . . _ _ . we H
A use: Hem» Hews Heme Heem nee: nee» _
.

 

i -- L

L 7] I man—mac: I I JM—uncwcmn ”as

D -- -lll- .. . .111 i .la .-.- - - i.iH..Iu|lrl.

 

m .1
_

 

 

 

 

AJ 0.5.50 m: 363“.” noun—.52. $18 :58” A33 n 33 2:— ><a3me 5.2.2.5. .. >532. wanes?“ 5—8

15

The ﬁrst half of the 19905’ were marked by the liberalization of capital movements in
August 1989. Thereafter Turkish Lira (TL) appreciated continously in real terms except
the two main depreciations. The ﬁrst one occured in 1991, the year of the Gulf crisis, and
the second one in 1994, the year Turkey experienced a severe ﬁnancial crisis. Before the
ﬁnancial crisis, Turkey promoted exports by means of a real depreciation of TL . This
policy changed after the crisis and Turkey promoted capital inﬂows by means of a real

appreciation of TL.

Public ﬁnances deteriorated markedly in 1992-1993 and political uncertainty
intensiﬁed. Combined with an open capital account, this led to a ﬁnancial crisis. The
main factors were unsustainable ﬁscal policy, the strong desire to keep domestic interest
rates below the market clearing levels, minimizing domestic debt ﬁnance and instead

relying upon monetization. See Ozatay, (1994).

Deterioration in ﬁscal balances and increased recourse to central bank ﬁnancing, lack
of ﬁscal discipline, and expansion of the credits extended to the public sector made it
impossible for the Central Bank of the Republic of Turkey (CBRT) to comply with the
monetary program in 1993. Reliance on foreign debt and the CBRT advances increased
by the end of 1993. The unsuitable environment for external borrowing necessitated
ﬁnancing of the deﬁcit through central bank resources. At the same time domestic
interest rates were administered such that the Treasury tried to pay less than the
equilibrium interest rate on government securities and several Treasury auctions were

cancelled. Moreover, an income tax on the holders of government bonds and bills was

16

introduced. The response of the private sector was not to purchase government securities
and, therefore, a funding crisis started. Excess liquidity and bank’s high foreign exchange
(FX) demand put upward pressure on the FX market. The turmoil in ﬁnancial markets
started in January 1994 and continued unabated through April, with the TL depreciating
sharply on a daily basis, despite high interest rates and frequent interventions by the
CBRT in the FX market. The US dollar appreciated by almost 70 percent in the ﬁrst three
months of 1994 against the TL. To prevent a further depreciation of the domestic
currency, the CBRT increased the overnight interest rates and intervened both in the
interbank money market and foreign exchange market. CBRT lost more than half of its
international reserves whereas the interbank money market rates increased to record
levels. The actions of the CBRT to stabilize the markets through open market operations
proved insufﬁcient and the rapid loss of reserves on a continuous basis called for more
radical and widespread measures. The ﬁrst devaluation was in January and the second
one was in April. See Figures 1.3 and 1.4. On April 5, when the CBRT stopped
intervening, the T L w as d evalued and the foreign e xchange m arket w as u nstable until
May 1994. International agencies decreased the rating of Turkey’s government debt and
three small b anks went bankrupt. H ence, the T urkish e conomy found itself i n a v ery

severe ﬁnancial crisis.

17

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

« A 4

3:223:

 

 

 

AJ Gem—w. mac» ”one on Hiram—V man—.amam me: .512. o». a 236572.. 58 93.5.. w: 532.52.. 5:

l8

 

 

A3

Hymn—=6 FA" 0.53.55!» Mano-6mg ”mace

wGQ

QGG

 

_. .8:

mac

 

 

 

 

 

L ace
_ we:
we:

_ :5

Nut-94

. m M H mm w W. W M W:

 

Ibo-94

 

AJ Dam—w. Cad—.3“... Ego—.8" ”was mun—Eu?“ ...:— ualca on n 232.52.. 33 33:“.— ue Gena—Ban... Sea

19

The CBRT promulgated several new regulations regarding ﬁnancial markets, and a
new monetary framework was prepared in line with the IMF stand-by agreement aﬂer the
crisis. In the medium term, in order to achieve price stability, a tight monetary policy was
put into effect. As stability reemerged in the markets, the treasury started to borrow from

the domestic markets. Foreign reserves rose strongly in the second half of 1994.

The second half of the 19905 was marked by monetary programs aiming at achieving
stability in the ﬁnancial markets. High inﬂation and an increase in the Public Sector
Borrowing Requirement (PSBR) kept the interest rates high during 1995-1999. Economic
growth was very strong except in 1999. The economy was characterized by high growth
rates and high inﬂation in this period. Following this strong demand-led growth, a sharp
decline was seen in 1999 due to the adverse effects of the earthquake in Turkey and the

Russian banking crisis. See Table 1.1.

The CBRT followed a monetary policy which focused on stabilizing the ﬁnancial
markets during 1995-1999. The CBRT announced publicly that its aim was to keep real
exchange rates stable. This aimed both at affecting agents‘ inﬂationary expectations and
to reducing the degree of uncertainty about the future of the economy. Under the
constraints of an open capital account, large ﬁscal deﬁcits and inertial inﬂation, CBRT
aimed at maintaining ﬁnancial market stability, which meant safeguarding
competitiveness and minimizing the burden imposed by high ﬁscal borrowing

requirements 0 n the financial sy stern. C BRT h as p repared a m onetary p rograrn e very

20

year since 1996. Although the programs were formulated with a certain inﬂation target in
mind, this did not quite mean that reducing inﬂation was the key objective. Instead the

overriding objective was real exchange rate stability and, ultimately, FX reserve strength.

According to the standby agreement signed by IMF in early 1995, the exchange rate
policy was based on the idea of utilizing exchange rates as a nominal anchor in curbing
inﬂation. The increase in the foreign exchange basket, deﬁned as 1.5 German marks and
1 US dollar, was targeted to increase by as much as the targeted monthly inﬂation rates.
The basket was revised as 1 US dollar and 0.77 Euro by the introduction of Euro in 1999.
Deviations from the targeted inﬂation rate were seen in certain periods and the targets for

the basket were adjusted accordingly.

During the managed ﬂoat exchange rate policy, most of the sales of US dollar were
seen in the last quarter of the year. The bank’s demand for foreign currency to close FX
open positions put upward pressure on FX markets. CBRT sold US dollars usually when
the German mark/US dollar parity rose, to achieve targets for the foreign exchange
basket, excess liquidity, speculative increase in FX demand due to adverse effects of
external crises like Russian and Argentine crises, an earthquake, and deterioration of
expectations about Treasury’s domestic debt ﬁnance. When there was stability in the
markets and capital inﬂow, or when the inﬂationary expectations decreased, CBRT

usually bought US dollars and the reserves increased.

21

 

 

 

a 1 -151,1.1515151-- 51:51 .- 1|-

ﬁmm—unc PM" mvcn ”an AHFam—VVAJ

 

WQHeeec .
_ _

_ mucccc

_
_s_eeee

.
unease

.

N H :25

1

 

1

 

 

l

 

 

 

l

 

_
L
“Hueeee

 

 

 

W Heeee M

3:335:33

 

 

 

 

 

 

AJ Gum—w. moon ”:3 on dram—v man—.535 =3 3: ~512— on N 55:93.. Sou «Sega—— 3 508552.. 58.

22

The economic environment was not good despite the stabilization efforts. Towards
the end of the 19905, deterioration in public ﬁnances, high levels of real interest rates and
inﬂation and the economic contraction necessiated the government to put a new medium
term program into affect in 2000. The Turkish government adopted a disinﬂation
program backed by the International Monetary Fund (IMF) at the beginning of 2000. This
program aimed at decreasing the inﬂation rate to single digits by the end of 2002 and had
three main speciﬁc targets. First, it imposed limits on CBRT’s balance sheet items such
that it limited monetary expansion only to changes in its net foreign asset position in the
balance sheet. The CBRT committed itself to a policy of no sterilization. Changes in
monetary base would be directly reﬂected changes in the net foreign assets of its balance
sheet. CBRT‘s liquidity creation was tied to foreign capital inﬂows. Within this
framework, the banks were sellers or buyers in the foreign exchange markets by
considering increases in the short-term interest rates. Second, pre-announced target
values of the currency basket, in other words a pre-announced calendar for the
depreciation rate in line with the targeted inﬂation, were applied. Third, there were
speciﬁc targets for primary surplus to ease the domestic debt ﬁnance. The CBRT

conﬁrmed that the exchange rate targets had been met by the end of 2000.

The disappearance of the foreign exchange risk caused nominal interest rates to
decline sharply. These developments in foreign exchange and interest rates led the
consumption of durable goods, production, investment expenditures, credits and imports

to rise and the Turkish economy, which had slowed down in 1998 and contracted in

23

1999, to enter a new period of growth in 2000. The monetary policy within the

framework of the “Disinﬂation Program’ ’ was carried out until February 2001.

The armounced increases in the basket at the beginning of the program were
according to the targeted but below the realized inﬂation rates. The decreases in nominal
interest rates were very sharp. This structure caused a deep deterioration in comparative
advantage and so the current account balance. Delays in the implementation of structural
reforms in the second half of 2000, and deviations from the privatization targets caused
loss of credibility of the program enhancing devaluation expectations. In line with the
monetary policy framework in which the CBRT’s liquidity creation was tied to foreign
capital inﬂows, the increasing capital outﬂows, due to devaluation expectations, resulted
in a rise in interbank money market interest rates in November 2000. Rising interest rates
caused deterioration in the ﬁnancial situation of banks with maturity mismatches in their
balance sheets, causing a lack of conﬁdence in the banking system. The ﬁnancial
system’s increasingly growing liquidity needs turned into ﬁnancial crisis and the CBRT
was left with no choice but to raise the liquidity it provided to the market. This situation
brought about a departure from the targets determined for the CBRT’s balance sheet
aggregates. Despite the fact that the crisis environment had been relatively subdued, the
maturity of both domestic and external funds lessened because of the rapid rise in the risk
premium, while the interest rates continued to stay at high levels relative to the currency
basket’s rate of crawl. Following the political turmoil on February 19, 2001, prior to the
Treasury’s auction, 3 great demand for foreign exchange emerged in the ﬁnancial

markets, but the CBRT tried to avoid the realization of the foreign exchange demand by

24

constraining the liquidity. This situation, however, resulted in a breakdown of the
payments system due to the need of state banks for excessive overnight liquidity. The
ﬂoating exchange rate regime was adopted on February 22, 2001 when the exchange rate
policy had become unsustainable under existing circumstances. A sudden capital outﬂow
caused a speculative attack on the foreign exchange market and, following the 22nd of
February, a new exchange rate system was put into effect. See Figures 1.6 and 1.7 and

1.8.

25

 

 

mam—=6 Faumﬁcn ”can Aﬁﬁam—VVE
Neccccc

 

Hmccccc 5;.

 

 

 

 

 

 

 

Hmcccac

unacccc

 

Huccccc -.

 

 

 

 

Haccccc

acaccc 1-

 

 

 

maaccc

 

 

 

 

 

 

 

bccaca

 

3:333:33

AJ Cam—u. ween 5.8 on dram—v man—5E.“ :5 E: .5..ch c.- u SEE—J; N25 33:”: n: bane—ace... N98.

 

26

 

 

 

 

‘+ \‘ -' I .111- .:-i ‘rjllli#
A

mam—:6 uh" 053.3%: 5.83m» ””263

 

:5

 

 

 

«a - , 15:5: ,;.14, 11 ---1, -, —

 

 

 

NHL. .17L3T z-
.3 i . E ¢ I

3 Sill! -

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

_
_
.
_ .3 4 .l 5111 1L1 _
_
_

 

 

 

 

 

 

 

 

 

 

 

AW? a

W W W. W, W
( W W W W W

Jun-00 ~
Jul-00 +

WWW W

|l.||lntull|'|.v |WIIL

 

 

AJ Dam—w. 0:35.: nus—.8» 5:8 man—5E5 :5 E: .512— 3. u has-5Q. N25 33:»: um Cane—van. ~25.

27

 

 

 

 

mam—:6 Fm "Cad—ﬁat" Han—.8” ””383
:c
:5 - ilii--,-i il i ,- l| ii: -, iii-i-,i-..
E. ii. .-i; , i -4 ,‘il--- i ii , iii , l Willi

 

 

 

 

 

ac .j E llil: i ii iliii - i i i
#6 iii 1 1 {iii ii- i, i iii i lliii-

a: i fll‘lrr . --i -. 1

ma - i ,- --ii

 

 

 

 

 

 

 

 

 

N:

WW W
WWWW

 

 

AJ Dam—w 053mm... 58.63 5:8 man—5:5“ =5 ...... .512— 3. u 32.2.. ~8— Snccmw 3 58352.. 38.

28

 

The banking sector problem in Turkey was basically a result of the mechanism
chosen to ﬁnance the very high public sector borrowing requirement. Firstly this led to an
increase in government debt instruments in the balance sheets of private banks and
caused a signiﬁcant deterioration in state-owned banks because of their accumulating
duty losses. There are two major interest-earning assets in the Turkish banking system-
commercial loans and government debt instruments. The share of the government debt
instruments portfolio in private commercial bank balance sheets was greater than the loan
portfolio. See Ozatay and Sak (2001). The quality of the government instruments
portfolio is directly related with the expectations regarding debt sustainability. This
feature increased the vulnerability of the banking system to concerns about the rollover of

the outstanding government debt instruments.

Yeldan (2001) argues the Turkish currency crisis did not originate due to the failure
of ﬁscal and monetary authorities in following the main targets of the program. The crisis
conditions emerged as a result of the increased fragility in the ﬁnancial system. Factors
such as weak prudential regulation over the banking sector and ﬁnance of the large
persistent ﬁscal deﬁcits were instrumental in the burst of the crisis. Ozatay and Sak
(2002) agree with him that a fragile banking system was the root cause of the crisis. The
banking system was highly vulnerable to capital outﬂows. The foreign exchange risk was
high in the period preceding the crisis. While the total open foreign exchange position of
the banking system was increasing, the ratio of liquid foreign exchange denominated
assets to total foreign exchange denominated liabilities was decreasing. Maturity

mismatch was an other problem. The liabilities were more of a short-term nature while

29

the maturities of assets were longer. Also the share of the bad loans to total loans was

very high.

The new program put into effect after the crisis aimed mainly at arranging the role of
the state banks and reducing their extensive short-term liabilities. The CBRT played a
key role in the operation of restructuring the state banks. The CBRT set its short-term
priorities so as to remove defects in the payments system, to make the ﬁnancial system
function again and to provide stability in ﬁnancial markets. To avoid the inﬂationary
effects, CBRT controlled base money by the implementation of monetary targeting.
Moreover, considering the absence of a nominal anchor due to the abandoning of the
exchange rate regime, money base targets were set to help the economic agents shape
their expectations after switch to a free ﬂoat regime. Short-term interest rates became the
most important monetary policy tool in dealing with liquidity control and inﬂationary

pressures during this period.

The exchange rate depreciated substantially and was volatile after the switch to the
ﬂoating rate regime. In the beginning of the ﬂoating exchange rate regime, the CBRT
declared that the exchange rate would be determined by market dynamics. The FX
interventions of the CBRT after the switch to the ﬂoating exchange rate regime was
directed towards damping the excessive volatility in the FX level without affecting its
long-run value. Moreover, a programmed auction system for FX sales was implemented
as of March 29, 2001. FX sales auctions were conducted daily until May 17, 2001.

Between May 17, 2001 and July 11, 2001, FX sales auctions conducted whenever

3O

required without pre-announcing the amount. After July 11, 2001 the monthly program of
FX sales auctions was pre-announced. These auctions were conducted in order to sterilize
the excess liquidity in the market caused by the use of the external ﬁnancing which was
provided by the IMF to make the Treasury’s domestic debt payments. The CBRT
sometimes intervened in the exchange rate markets in order to smooth the ﬂuctuations
which emerged in the case of negative external developments and domestic political
problems. However, beginning in August, non-programmed interventions were reduced
to negligible levels, and as of September, program based auctions were carried out on the
daily basis. After November 30, 2001, FX sales auctions were conducted when they are

required.

The CBRT pursued foreign exchange buying auctions in order to absorb the excess
supply of foreign exchange that was accumulated by reverse currency substitution and the
surplus in the balance of payments in the ﬁrst half of the 2002. These operations did not
aim at distorting the long-run trend in exchange rates. Moreover, the CBRT did not try to
establish any speciﬁc exchange rate level. Foreign exchange auctions were temporarily
suspended as of July due to the volatilities of exchange rates, which occurred as a result
of an increase in the perception of political uncertainty. Similar to 2002, in 2003 the
CBRT intervened in the foreign exchange markets when there was high volatility and the

operations did not aim to distort the long-run trend in exchange rates.

Turkish economy was quite unstable through 19903. Because of the two ﬁnancial

crisis in 19905, it was realized that, without structural reforms, disinﬂation programs

31

ll

have no chance to decrease the inﬂation so priority was given to structural reforms
regarding the arrangements of economic institutions at the beginnings of 2000s and
disinﬂation program has found more support after the crisis. Tight monetary and ﬁscal
policies helped to achieve ﬁnancial stability, which created a suitable environment for

economic growth.

1.3 Data

In this study, the data is provided by the CBRT. The data sample consists of both
daily spot bid rates, interbank overnight interest rates and daily intervention variables
from November 1993 through December 2003. Intervention values are in millions of US
dollars. Exchange rate policies are different throughout the period. The analysis has been
done separately for managed ﬂoat and free ﬂoat exchange rate policy regimes,
respectively, including and excluding the crisis periods. So we take several sub-samples
of the d ata. T he first 5 ub-sample h as the d aily amount 0 f 11 et d ollar p urchases (sales),
daily spot bid rates and interest rates date from November 1993 through December 1999.
This sub-sample covers the ﬁrst ﬁnancial crisis in April 1994 and exchange rate policy
was managed ﬂoat. Then the crisis period is excluded and only the period 1995-1999 is
analyzed. This enables us to investigate the effect of intervention in both a crisis
environment and without it. T he third sub-sample d ates from 2 000 through D ecember
2003. This period covers both crawling peg and free ﬂoat regimes and include the second
ﬁnancial crisis. Then we exclude the ﬁnancial crisis and estimate only the free ﬂoat
regime from February 27, 2001 through 2003. The daily spot returns for whole period is

shown in Figure 1.9. The graphs of the spot returns and interest rates are drawn for

32

certain sub-samples in order to show the data better. The ﬁrst sub-period consists of
November 1993-1994 at which time exchange rate policy was managed ﬂoat. In this
period the ﬁrst ﬁnancial crisis happened. Daily spot returns and short-term interest rates
are seen in Figures 1.10 and 1.11. The period 1995-1999 is marked by IMF- standby
agreements. Exchange rate policy was managed ﬂoat and the exchange rate basket was
pegged to forecast inﬂation. Daily spot returns and short-term interest rates are seen in

Figures 1.12 and 1.13.

33

 

 

a:

 

we - i ili l i

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

m
N

Nov-93
Nov-94
i Nov-95
i Nov-99
Nov-00
Nov-01
Nov-02

 

 

 

AJ 6:3. mi: ”25.: en dram—v mun—5:3” :5 E: 1012. 2. N 2363—5... new... 2.3.5.. 3 Ugo—Baa... 33.

34

 

 

mam—=6 PS" mug ”3:35 $55253

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

, um
__ .3 _ ii -.i iii
ail; _ I i
3 iii [
5i _ i
M _ _
.. \__>.,.i
-m-i, 1 _
-3 I i
Lm
M M. M M M M M. M. m M M. m M M.

 

 

 

 

3 Us”? mg» ”02:... on dram—u mun—5:5 :5 E: .312— 3 N 238552.. 33 33:»: mo Dona—awn... 33.

35

 

 

mac

 

 

.25

 

ac:

 

 

 

 

 

 

 

 

mac

 

 

 

 

£6:

 

 

 

 

 

Mac

 

 

 

 

 

Na:

 

 

 

 

 

 

:5

 

 

 

 

 

 

 

 

AJ Dam—v. C5355 3:26.: 5:8 man—.5?" E: .512— 3. — 235552.. 33 93:“.— uc 582—52.. 5:

36

 

 

mam—=6 HAN“ m—x: ”2:35 AHFEmUVS

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

.

.
.. i i - ,
..1 iii i “
N L i ii - _ i - i i ii M
H p — p -—-i b _
_. , m
.— _ i - - .
bit: I- i I I Iiii ». 1 ii I .
TM ‘
AMZMMZMMZMMMMMMMMMMM
mmMmmMMmmmMmmmMmmMMm.

 

 

 

AJ 59:! mean ”25.: a. dram—u mun—~53" :8 3:12.25 3. N .35:qu 5cm 93.5.. 3 5325—3... 58.

37

 

 

 

Hui—=6 rum" 048-5539 7:9.an ”means

 

Mm:

NGG +l||€+eisr; :s- : . ‘i;:‘\“-!

 

 

 

# .mc 11.? ix

 

 

 

 

u :51

 

m6 1

 

 

 

 

 

ﬂ ..

:MZWZZMZEZWZ

 

L—..__ .__.__~_ _._.—... ___~.. ._w.—_

 

3593 0335—: man—.8" 5:8 man—5:5“ :5 E: “512. on N amass. Gem ...—.25: 3 528532.. 53.

38

The CBRT changed the exchange rate policy and adopted crawling peg regime in
2000. However, this policy was not sustainable aﬁer the second ﬁnancial crisis and free
ﬂoat regime is adopted in 2001. Daily spot returns, and short-term interest rates are seen
in Figures 1.14 and 1.15, 1.16. During the crisis period interest rates increased to very
high levels. It is difﬁcult to show interest rates in one graph so we separate the period and

graph the interest rates as before and aﬁer crisis.

39

 

It lvillliil)

_ Haw—5.0 HLAu maven Won—=5 AH‘FEWUVAJ _
_ a: n s

 

 

Na ., --- ill ,4, llli-

 

 

 

 

 

 

 

 

 

 

AJ Cum—v. mac. ”SE... 3. dram—v man—5:5“ :5 E: .512. on u ass—SQ. ~25 93:“: M: 58253... 33.

4O

 

 

:5
m5

 

 

 

 

 

a:
m:
as
w:

H:

wmmwww mm m N

do Dun—w. 033—5.: min—.8. 5:8 man—5:3” 9a ...... 12.3.. on u SEE—Q. Nee: 933m.— .3 Cane—.0... was.

 

 

41

 

 

 

 

 

mam—=6 HLQOJGHBmm—un Mano—6m» Enema; .
_ .
.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

‘ iii ‘ l I‘ .IliL

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

a .s a ll Will, :IWWWI‘JJJJ

q 4 a a a a d _ Jl a a

3:333:33

 

 

 

 

 

diva—vs GEE—mug man—.8. 258m man—5:5“ =5 «5:312. en a 326... N93 Ewosmr 3 6305—5... 38.

42

1.4 Model and Estimation Output

In this section, we aim to test the direct effect of intervention on the level and
volatility of spot returns. In the literature, there is mixed evidence about the effectiveness
of intervention. 80 this analysis is a contribution to the literature such that we are able to
analyse the effect of intervention in Turkey. Turkish economy is a small economy where
the central bank is an important player in the foreign exchange market. So we expect the
foreign exchange intervention to be effective. Also, this study enables us to investigate
the effect of intervention under different exchange rate policy regimes and crisis

environment. Intervention is proposed to affect the mean and volatility of the spot rate.

The model in this study is built around the standard Martingale model with time
dependent conditional heteroscedasticity. Following Bollerslev (1986) and Baillie and
Bollerslev (1989), the conditional variance is modeled as a linear Generalized
Autoregressive Conditional Heteroscedasticity (GARCH) process (see equation 1.3).

Bollerslev (1986) introduced the GARCH (1,1) process, which extends the ARCH model
to make 0,2 a function of lagged values of 0,2 as well as the lagged values of 8,2

Bollerslev (1986) required all the coefﬁcients to be positive to ensure that the conditional
variance is n ever negative. The q uasi- m aximum likelihood e stimation is u sed. In this
study, the aim is to investigate the effect of central bank intervention on the foreign
exchange markets. We used a linear GARCH (1,1) model following Baillie and

Osterberg ( 1997b) .

43

4 n
b . .
AlnSt = b0 +b1USt_2 +b2US,_25 +b3z,_2 +'21di +2111. +5, (1.1)

l:

where A In S, =100[In(St)— In (s,_1 )]
a,=z,a, with z, is iid (0,1) (1.2)

0,2 = a) + aat_12 + ,60',_12 + leSt_2b + 72USt_ZS (1.3)

S is the level of the spot exchange rate in terms of the number of TL per one USS.

l
The variable AlnS: is the return on the exchange rate between 15.30 pm on day H and
15.30 pm on day t. The explanatory variables includes the intervention variables and

short term interest rates, d, corresponds to the daily dummies and j, corresponds to the

crisis dummies. Crisis dummies are used in the conditional mean to take control of the
devaluations during the crisis period. j] and j 2 corresponds to the dummies to take control
for the devaluations in January 1994 and April 1994, respectively, while j3 and L;
corresponds to the dummies to take control for the following two days after the

devaluation on April,5 1994.

The intervention variables US b and US ‘ are purchases and sales of US dollars.
Empirical results of estimating this model have generally been poor, probably because of
the policy endogeneity issue, where a purchase of $ is found to be associated with a $
depreciation due to “leaning against the wind” phenomenon. See Baillie and Osterberg

(1997) and Humpage (1988). The intervention variables are lagged to ensure they are

44

pre-determined to a avoid the possible simultaneity bias. The intervention variables are
the logarithms of net purchases or sales of intervention ending at the close of the markets

on day t-2. If no intervention occurs, then the intervention variable is set to zero. The

short-term interest rate (i) is also lagged as t-2 and is the logarithm of (1+ f—gé) where

on, is the overnight interest rate. The empirical results of the estimated models are given

in Table 1.2 and Table 1.5. Asymptotic standard errors are in the second column with the

corresponding parameter estimates. The statistics m3 and m 4 are the sample skewness

and kurtosis of the standardized residuals. Q20 and Q202 are the Ljung-Box test

statistics for autocorrelation and ARCH effects, and T is the sample size.

The model is estimated for different sub-periods. First, the managed ﬂoat period
including the ﬁnancial crisis, covers November 1993 through December 1999. The
estimation output is shown in the Table 1.2.During the managed ﬂoat period, intervention
is sterilized and Emir et al (2000) showed that there was a low degree 0 f sterilization

before the 1994 ﬁnancial crisis and a high degree of sterilization during 1995-1999.

The crisis dummy variables in the mean equation are statistically signiﬁcant and their
magnitude is large. There is also a signiﬁcant Monday effect and this is most likely due
to the calculation of daily target values for the basket according to the projected inﬂation
in a m anaged float regime. T he p rojected m onthly increase in the b asket is linearized
through the whole month, including the weekends, so the change between Friday and
Monday includes target values for Saturday and Sunday as well. Also, most of the

announcements w ere (1 one towards o r o n the w eekends w hen the m arkets are closed.

45

Most of the responses to the news are reﬂected on Monday when the markets are open.
The most important ﬁnding regarding to results for managed ﬂoat period is that
intervention h as b een unable to inﬂuence the c hange in the exchange rate. D uring the
managed ﬂoat regime, the aim is to achieve the targets for the basket which consists of
US dollar and German mark (euro after 1999). After the crisis, the turmoil in ﬁnancial
markets continued unabated through several months, with the TL depreciating sharply on
a daily basis, despite high interest rates and frequent interventions by the CBRT in the F X
market. The empirical results suggests that the short term interest rate has been able to
inﬂuence the exchange rate such that increase in interest rates appreciates the US dollar.
A 2002 study by Gumus supports our ﬁndings as it reports that raising interest rates had
a signiﬁcant effect on depreciating the nominal exchange rates. There is persistent
volatility in the conditional variance. The sum of the coefﬁcients of squared residual and
the lagged value of variance is close to 1. The kurtosis of the standardized residuals is
high. Skewness is also high and positive. There is also evidence that intervention have
no effect on the volatility of the daily spot returns. During this period CBRT aimed at

reducing the volatility in real exchange rates.

46

Table 1.2: Estimation of Intervention Mode

Conditional Mean Parameters

l (a)

 

(I)
a

13
"Yr

'Yz
Skewness

Kurtosis
on2
020
'1‘

0.2877
0.3249
0.0198
0.0220
0.0327

12.2960 "
35.8032
25.2516
43.3175 °
Conditional Variance Parameters

0.0208
0.2192
0.7214
-0.0021

-0.0019

Coeﬁicient
-0.0049

-0.0003
-0.0062

.0.

.0.

Standard Error

1.34
17.55
30.76
15.28

1558

0.0842
0.0070
0.0070
0.1630
0.03 86
0.0291
0.0346
0.0310
0.2430
2.6724
2.61 18
2.8799

0.0092
0.0687
0.0865
0.0019
0.0019

1" Full period of 1 November, 1993 through 31 December, 1999.
(‘0 denotes 10% signiﬁcance level
('1') denotes 5% signiﬁcance level
(...) denotes 1% signiﬁcance level

AlnSt =b0 +b1US,_2b +b2USt—ZS +b3i,_

2

2+.

1

at = a) +a8t_12 + 60,42 + 71US1_2b + yzUSt_23

5t =21“:

47

with z, is iid (0,1)

4 n
2 di +2 ji +8,
=1 i=1

Table 1.3: Estimation of Intervention Model ‘a’

Conditional Mean Parameters

 

Coefficent Standard Error

b0 -0.0555 0.0753
bl 0.0007 0.0061
b; -0.0035 0.0063
b, 0.3884 0.1375
d1 0.3267 0.0332
d2 0.0022 0.0257
113 0.0039 0.0278
d4 0.0028 0.0274

Conditional Variance Parameters

 

to 0.0134 0.0045

01 0.1641 ” 0.0675

[3 0.7267 0.0635

7, 0.0002 0.0012

72 0.0001 0.0013
Skewness 0.48
Kurtosis 9.92
Q20 31.86
onz 15.02
T 1259

 

('0 Full period of 1 November, 1993 through 31 December, 1999.
(*) denotes 10% signiﬁcance level

0") denotes 5% signiﬁcance level

(...) denotes 1% signiﬁcance level

4 n '
2 +§1di+i§11i+£t

AlnS, =00 +b1USt_2b +02US,_25 +b3it_
I

0,2 = a) + (16,42 + ﬂat_12 +y1USt_2b + 72USt_23

5t=zt0t with Zt is iid (0,1)

48

The intervention models were also estimated for the period 1995-1999, which
excludes the crisis period. The empirical ﬁndings shown in Table 1.3 for period 1995-

1999 are similar to the results for 1993-1999.

Intervention has been unable to inﬂuence the change in the exchange rate. The signs
are as expected but not signiﬁcant and there is a signiﬁcant Monday effect. Overnight
interest rates are signiﬁcant but not of the expected sign. There is a tendency that the
increase in o vemight rates is a ssociated with a subsequent d ollar appreciation. D uring
this period inﬂation is high and agents have inﬂationary expectations. They already know
that the US dollar will appreciate as long as inﬂation is high. The basket which consists
of the US dollar and TL is pegged to inﬂation. High interest rates indicate high inﬂation
and result in a strong FX demand. The intervention variables have no effect on the

volatility during this period.

The empirical results of the models estimated for the period 2000 through 2003 are
quite different from the models for the m anaged float regime. During the D isinﬂation
Program covering the period January 1, 2000-February 21, 2001, intervention is not
sterilized but after a switch to a ﬁ'ee ﬂoat regime, intervention is fully sterilized. This
period includes the second ﬁnancial crisis. The empirical results are shown in Table 1.4.
Linear integrated GARCH (1,1) model is estimated. The coefﬁcients of intervention are
signiﬁcant. Purchases of US dollars are associated with subsequent dollar appreciation.
This is most probably due to that the central bank intervenes (purchases) with the wind
during the crawling peg period. Sales of US dollars is followed by the appreciation of the

US dollar. This effect is sometimes known as the leaning against the wind phenomenon

49

and suggests the possibility of policy endogeneity with central banks selling dollars
because the dollar is appreciating. The TL depreciated substantially and displayed a
highly volatile pattern in the ﬁrst months after the switch to the free ﬂoat regime. The
CBRT conducted regular FX sales auctions to smooth excessive short-run exchange rate
ﬂuctuations without affecting the long-run equilibrium level of exchange rates and to
perform FX sales in a more transparent manner in the ﬁrst half of 2001. Contrary to a
managed ﬂoat period, increase in interest rates decrease the spot returns as expected. The

crisis dummies are also statistically signiﬁcant. jl corresponds to the devaluation day,
February 22, 2001 and j2 and j3 corresponds to the following two days after the crisis,

respectively.

We exclude the crisis period and estimate the same models for the free ﬂoat period for
February 27, 2001 through December 31, 2003. The empirical ﬁndings are shown in
Table 1.5. After the switch to a free ﬂoat regime, CBRT declared that the exchange rate
would be determined by market dynamics and it would not intervene in the exchange rate
markets except in cases where the exchange rate displayed an instantaneous and highly
volatile pattern. Intervention is ﬁilly sterilized and, more importantly, announced to the
market. The empirical ﬁndings for the free ﬂoat regime have similarities with the ones
for managed ﬂoat regime. Intervention has been unable to inﬂuence the change in
exchange returns. Contrary to Domac and Mendoza (2004), although the central bank
advocated the use of intervention to reduce exchange rate volatility, sales of US dollars
may have actually Granger caused an increase in volatility and this result may reﬂect that

intervention, by providing more information to market participants, which can induce

50

increased volatility as agents adjust their positions. Similar to intervention, overnight
interest rates are also not effective during this period. The sign is as expected but not

signiﬁcant.

51

Table 1.4: Estimation of Intervention Model 0"

 

Conditional Mean Parameters

Coefﬁcient Standard Error

b0 -0.0694 0.0723
bl 0.0390 0.0112
b2 0.0306 0.0111
b3 -0.2900 ” 0.1238
11, 0.1804 0.0544
d2 0.0940 0.0603
113 0.1400 ” 0.0573
d4 0.0990 ’ 0.0569
j1 33.0800 0.8556
j2 10.9346 1.0750
j; 43.9610 1.1828

Conditional Variance Parameters

to 0.0288 " 0.0183

[3 0.2115 ‘ 0.0954

7, -0.0055 0.0034

72 -0.0031 0.0028
Skewness 0.608
Kurtosis 5.538
Q20 32.68
onz 26.98
T 1004

 

(a) Full period of 1 November, 1993 through 31 December, 1999.
(*) denotes 10% signiﬁcance level

('1') denotes 5% signiﬁcance level

1*") denotes 1% signiﬁcance level

4 n_
2di+Zji+5t
' i=1

AlnSt =b0 +b1US¢_2b mgr/5,45 +b3it_ 1
1:

2+

O'tz = a) +a8t_12 +ﬂ0't_12 +}’1US1_2b +72USt_28

8t=2t0t With zt is iid (0,1)

52

Table 1.5: Estimation of Intervention Model (”0

 

Conditional Mean Parameters
Coefﬁcent Standard Error

b0 -O.2198 0.1987
bl 0.0202 0.0203
b, 0.0295 0.0497
b, 0.0998 0.5049
d1 0.1878” 0.0371
d2 0.104 0.1071
6, 0.2094“ 0.104
d4 0.0768 0.0962
Conditional Variance Parameters
to 00834 0.0397
01 0.2954” 0.1 133
p 0.6105‘” 0.1129
7, 0.0019 0.0097
72 0-192" 0.0838
Skewness 0.9
Kurtosis 5.26
Q20 29.4
Q202 15.1 1
T 712

 

('0 Full period or 1 November, 1993 through 31 December, 1999.
0’) denotes 10% signiﬁcance level

(...) denotes 5% signiﬁcance level

(...) denotes 1% signiﬁcance level

4 n
2 +§1di +311}. +8,

AlnS, =b0 +b1USt_2b +b2US,_2‘ +b3it_
1_

0’2 = a) +a£t_12 +ﬂ0't_12 +71USt_2b +72USt_2S

€t=ZtUt With zt is iid (0,1)

1.5 Conclusion

In this section we investigated the effect of intervention on both the level and
volatility of spot returns in Turkey. The analysis covers the period November 1, 1993-
December 31, 2003. It is a common belief among policy-makers that central bank
intervention is effective. However, the evidence on the effectiveness of intervention is
mixed. In general, the evidence is disposed towards ineffectiveness. We also ﬁnd
evidence that central bank intervention was not effective in Turkish foreign exchange
market during the period November 1, 1993-December 31, 2003. We estimate a linear
GARCH ( 1,1) model for different sub-periods including and excluding the crisis periods.
The results show that intervention has been unable to affect the level of spot returns both
in managed and ﬂee ﬂoat regimes except during the sub-period January 1, 2000-February
21, 2001. While intervention has no effect on volatility in managed ﬂoat regime, it
increases the volatility in a ﬂee ﬂoat regime. With respect to effect of short-term interest
rate, an increase in overnight interest rates appreciates US dollar in a managed ﬂoat

regime but has no effect in a ﬂee ﬂoat regime.

54

CHAPTER 2

What Motivates Central Bank Intervention?

2.1. Introduction

In the ﬁrst chapter, we ﬁnd evidence that central bank intervention was not effective
in Turkish foreign exchange market during the period November 1993-December 2003.
In this chapter we ask the question what motivates central bank intervention and analyze

the estimation of intervention reaction ﬁmction.

A proposed relationship between the amounts of intervention and the various
characteristics of a disorderly foreign exchange market is called a reaction function.
Since no speciﬁc measures of the disorderly market conditions are given by either the
central banks or by economic theories, economists often rely on econometric
methodologies to test if certain characteristics of the market are closely related to the
interventions. One of the main challenges in specifying a reaction function is the fact
that the intervention variable has a zero value for the majority of the observations in a
sample while the explanatory variables are not zero. This implies a non-linear
relationship because the observed quantities of intervention do not increase or decrease
approximately in proportion to the level of the explanatory variables. One way to proceed

is to approximate this potential non-linear relationship with a linear model. Eij fﬁnger and

55

Gruijters (1991), Ito (2002), Rogers and Siklos (2003) used simple OLS models. An
other way is to model the probability, rather than the quantity of intervention, uses a
probit approach, as in Baillie and Osterberg (1997), Dominquez (1998), Kim and Sheen
(2002) and McKenzie (2004). Recent studies use ordered probit as in F renkel, Pierdzioch
and Stadtrnann (2003), and Ito and Yabu (2004). Different ﬂom the probit model,
Frenkel and Stadtrnann (2001) use the logit model. If one is interested in the quantity,
rather than the probability of intervention, an appropriate model may be a Tobit model.
Almekinders and Eij fﬁnger (1994) and Humpage (1999) follow this approach. However,
a Tobit approach takes either a buying intervention or a selling intervention one at a time
but not both as the dependent variable. If one wants to explain both types of intervention
simultaneously , then the ﬂiction model may be an appropriate speciﬁcation for the
reaction function as in Almekinders and Eijfﬁnger (1996), Kim and Sheen (2002) and

Sheen (2002).

In the ﬁrst chapter we estimate the effect of intervention on the spot returns using
linear GARCH(1,1) model for different sub-samples including and excluding the crisis
period. In this chapter, we only take the sub-samples of managed ﬂoat and ﬂee ﬂoat
period excluding the crisis period. During the crisis period, the motivations for
intervention may be quite different ﬂom normal times. We ﬁrst used the probit model to
assess the probability of intervention. We then use a tobit model and compared the results
and found that the main assumptions of the tobit model of homoscedasticity and
normality are likely to be invalid. Hence we used the censored least absolute deviation

method, as suggested by Powell (1984) to get asymtotically consistent estimators. Section

56

2.2 gives a description of the measures of market disorder and the data. Section 2.3
brieﬂy mentions the probit model and gives the estimation output with respect to probit
model. Then Section 2.4 gives a description of the tobit model and the censored least
absolute deviation method in the context of intervention . The empirical results with
respect to these models, are also given in section 2.4. Finally, section 2.5 gives a brief

conclusion.

2.2. Modeling Intervention Behaviour

We propose that a central bank intervenes with the objective of minimizing
disorderliness over time in the foreign exchange market for its currency, depending on its
perception of the effectiveness of intervention. First, they might wish to reduce
disorderliness by returning the exchange rate to what they perceive to be the appropriate
trend to be. With a very long horizon, purchasing power parity considerations might drive
intervention behaviour. See Dominquez and Frankel (1993) and Frenkel and Stadtrnann
(2001). Secondly a central bank may be concerned about disorderly conditions in foreign
exchange markets that might show up as excessive ﬂuctuations in exchange rates through
higher volatility, due to higher levels of uncertainty and trading. They may intervene to
cahn the market by trying to reduce uncertainty. This uncertainty may be measured by

the conditional volatility of the daily change in the exchange rate.

Most empirical studies include the deviation of exchange rate ﬂom some targets

levels and a measure of volatility of the exchange rates. There is no clear answer to the

57

appropriate measures of the disorderly market conditions and it is rather difﬁcult to reach
this information. The target level for the exchange rate is thought to present past levels of
the exchange rate. This is not to say that the exchange is considered to be at a desirable
level in previous days. It merely allows one to test whether the central bank
systematically ‘leaned against the wind’ and tried to smooth deviations ﬂom the m-days
moving average of the exchange rate. Central banks appear to intervene whenever

current exchange rate movements deviate signiﬁcantly ﬂom a trend.

The deviation measure (devt) is modeled in (2.1) as the deviation of the current

exchange rate ﬂom a moving average exchange rate.
m 1 m
devt =10q1n(St)—-In(— 2 St —i)] (2.1)
m i=1

Different studies include different length of a representative moving average.
Almekinders and Eij fﬁnger (1994,1996) include the deviation of DM/USD rate ﬂom the
seven-day moving average . Humpage (1999) uses ten-day moving average. Frenkel and
Stadtrnann (2001) and Frenkel, Pierdzioch and Stadtrnann (2003) use a 25- day moving
average as a short run target level of exchange rate, and purchasing power parity as a
long run target. Kim and Sheen (2002), Neely (1998) and Le Baron (1999) use a ISO—day

moving average rate and claim it is the common choice among market traders.

58

As for the volatility measure, the dominant choice seems to be the conditional
variance of the log return of the exchange rate estimated with a GARCH(1,1) model. The

conditional variance is estimated using the following model:

100[In(St) — In(St_1)] = ,u + at (2.2a)
0,2 = a) + 016,42 + 60,42 (2.20)

where z, is i.i.d (0,1). There is considerable evidence that shows central banks respond to

deviations of the spot rate ﬂom some target level by leaning against the wind, and to
exchange rate volatility, by market calming down. Ahnekinders and Eijfﬁnger
(1994,1996), Frenkel and Stadtmann (2001), and Kim and Sheen (2002) ﬁnd evidence of
leaning against the wind and of market calming. They show that both target deviations
and volatility matter. Baillie and Osterberg (1997), ﬁnd that while a GARCH measure of
the deviations of conditional volatility ﬂom unconditional volatility has no effect on
intervention, (1 eviations o f the spot rates from the targets d o m atter. A s in B aillie and
Osterberg (1997), Ito and Yabu (2004) show that deviations ﬂom the target motivates
intervention. Dominquez (1998) and McKenzie (2004) ﬁnd no signiﬁcant intervention

response to volatility of exchange rate movements.

It is difﬁcult to reach the appropriate measures for disorderly market conditions.
According to the standby agreement signed by the IMF in early 1995, the exchange rate

policy was based on the idea of utilizing exchange rates as a nominal anchor in curbing

59

inﬂation. The increase in the foreign exchange basket, deﬁned as 1.5 German marks and
1 US dollar, was targeted to increase by as much as the targeted monthly inﬂation rates.
The basket was revised as 1 US dollar and 0.77 Euro by the introduction of the Euro in
1999. Deviations ﬂom the targeted inﬂation rate were seen in certain periods and the
targets for the basket were adjusted accordingly. Target values of the data for this period
are unavailable. The FX interventions of the CBRT, after the switch to the ﬂoating
exchange rate regime, were directed towards damping the excessive volatility in the FX
level without affecting its long-run value. In this analysis, we use several lengths of a
representative moving average values such as 5, 25, 50, 125 days for both the managed
and ﬂee ﬂoat periods. This enables us to see the response of central bank to short and

long-term deviations.

2.3.1. Probit Model

Econometric modeling of the daily intervention series has some practical challenges.
The dependent variable, e. g. the intervention series, is discontinous and so modeling it
using standard regression techniques is inappropriate. The intervention variable takes
zero values for the majority of the observations in a sample while the explanatory
variables are not zero. This implies a non-linear relationship because the observed
quantities do not increase or decrease approximately in proportion to the levels of the
explanatory variables. We generate a binary choice dependent variable corresponding to
intervention/no intervention outcomes for each of the two types of interventions and

model the probability of each type of intervention. As an initial approach, we follow

60

 

Baillie and Osterberg (1997), Dominquez (1998), and Kim and Sheen (2002) and adopt a
probit model and estimate the probability of positive and negative interventions of the

CBRT‘s foreign exchange market interventions.

In a binary response model, the aim is to ﬁnd the response probability

P(y=1|x)=P(y=l|x,,x2,....x,) (2.3)
where x denotes the full set of explanatory variables. Wooldridge (2000) mentions that
the linear probability model is simple to estimate but has some important disadvantages-
the ﬁtted probabilities can be less than zero or greater than one and the partial effect of
any explanatory variable is constant. To avoid these limitations, a class of binary
response models is proposed and, among the nonlinear functions suggested, probit is the

most commonly used model. Because of the nonlinear nature of E ( y | x) , a maximum

likelihood estimation is used (MLE). The MLE is consistent , asymptotically normal and
efﬁcient. Likelihood ratio statistics (LR) commonly are used to test multiple restrictions

in probit models.

We deﬁne US,bas the amount of dollar purchases vis-a-vis TL by the CBRT in time

periodt. Similarlywe deﬁne US,’ asthe amountofdollar sales visa visTLintime

period t.
Itb =1, (18,150 and Itb = 0, otherwise (2.4a)
1,5 =1, US,S)O and 1,3 = 0, otherwise (2.4b)

61

Then the following two equations are estimated separately by using a Maximum

Likelihood Estimation

Itb = 71b + yzbdevt_1(m) + 73b0't_12 + gtb (2.5a)
and

s _ s s (m) s 2 s
It —y1 + 72 devt_1 + 73 0',_1 + at (2.5b)

It is important to mention that we also include the lagged value of the dependent
variable among the regressors. A 2004 study by De Jong and Woutersen considers

dynamic time series binary choice model in detail.

In this study, the data is provided by the Central Bank of Turkey. The data sample
consists of both daily spot bid rates, and daily intervention variables ﬂom November 1,
1993 through December 31, 2003. Intervention values are in millions of US dollars.
Exchange rate policies are different through out the period. The analysis has been done
separately for managed ﬂoat and ﬂee ﬂoat regimes excluding the crisis periods. So two
subsamples of the data are taken and the ﬁrst has the daily amount of net dollar purchases
(sales) and daily spot bid rates date ﬂom January 2, 1995 through December 31, 1999
when the exchange rate policy was managed ﬂoat. We have the estimations for only the

ﬂee ﬂoat regime which dates ﬂom February 27, 2001 through December 31, 2003. The

data used in the estimations will have different starting values because the devH'" are

calculated starting ﬂom January 2, 1995 (for managed ﬂoat) and ﬂom February 27, 2001

62

25

(for ﬂee ﬂoat). For example, when we estimate the reaction ﬁrnction using devH and

volatility, the estimation sample dates ﬂom February, 7 1995 to December 31, 1999 for

the managed ﬂoat period and dates ﬂom April, 10 2001 to December 31, 2003 for the

5

ﬂee ﬂoat period. When we use devH 0 and volatility, the estimation sample dates ﬂom

March 15, 1995 to December 31, 1999 for the managed ﬂoat, and May, 16 2001 to

December 31, 2003 for the ﬂee ﬂoat period.

2.3.2 Estimation Results

The estimation output for the managed ﬂoat regime is shown in table 2.1. We take all
the representative values of deviation ﬂom the m-day moving average and estimate the
reaction function and include only the signiﬁcant ones. For the managed ﬂoat period, we
ﬁnd that deviation ﬂom both ﬁve and 25 day moving average increases the probability
of sale while deviation ﬂom the 50 day moving average and volatility decreases the
probability of sale. The probability of purchase of US dollars is increased by the
deviation ﬂom a 50 day moving average and is decreased by the deviation ﬂom a 25 day
moving average. The probability of purchase is not determined by the conditional
volatility. Lagged values of intervention increase the probability of intervention next day.
The empirical ﬁnding is consistent with the monetary policy of the central bank during

this period. The bank aims to achieve stability in real exchange rates.

63

 

Table 2.1: Estimation of Probit Intervention Model (‘0

Purchases of US Dollars Sales of US Dollars
Coefﬁcient Standard En'or Coefﬁcient Standard Error

C 0.75 0.120 0.90 0.12

DEV...’ -011 0.09 0.19 ' 0.11

om,” -0.28 0.100 0.28 0.11

DEVHS" 0.16 0.060 0.13 " 0.06

o,,’ 0.27 0.26 -0.56 ‘ 0.32

1,,” 0.980 0.080

14’ 0.94 0.08
McFadden 112 0,12 0.13
LRstatistic(5 01) 199 203
1"” 1209 1209

('0 Full period of 15 March, 1995 through 31 December, 1999.
(b) T denotes sample size

0') denotes 10% signiﬁcance level

(ﬂ) denotes 5% signiﬁcance level

(...) denotes 1% signiﬁcance level

b b

2
1t =71 +72bdth—1(m)+73b0't—1 +€tb

where 1,” = 1, US,”>0 with 1,” = 0,0therwise
and

2
Its = 713 + yzsdevt_1(m) + 935034 + as

where 1,3 =1, UStS )0 and Its = 0,0therwise

64

The estimation output for the free ﬂoat regime has differences from the managed
ﬂoat. During the ﬂee ﬂoat regime, the CBRT announced that the aim is not to affect the
level of the exchange rate. The aim is to smooth the excess volatility. Interventions do not
affect the long-run equilibrium of the exchange rate. Tables 2.2 gives the estimation
output for the probit analysis during the ﬂee ﬂoat regime. We ﬁnd that the probability of
sale of US dollars is increased by the current volatility of the exchange rate and the
deviation ﬂom both the ﬁve and 125-day moving average. Deviations ﬂom the ﬁve-day
and 50-day moving average increase the probability of purchase while deviations ﬂom
125 day moving average decrease the probability of purchases. A lagged value of
intervention increases the probability of intervention next day. These results indicate that
the central bank is more responsive to longer term deviations and volatility during the

ﬂee ﬂoat regime.

65

—
Table 2.2: Estimation of Probit Intervention Model 0')

Purchases of US Dollars Sales of US Dollars
Coefﬁcient Standard Error Coefficient Standard Error

C -111 0.13 4.15 1.04

DEVHS 0.13 ” 0.06 0.22 ” 0.10

111sz50 0.05 ‘ 0.02 -024 ” 0.11

Driv,,,'25 -011 0.01 0.23 ” 0.10

of 0.02 0.11 0.44 ” 0.20

1,," 1.45 0.15

1.: 2.58 0.35
McFadden R2 0.45 0.820
LR statistic(5 d1) 339 335
1“” 587 587

m Full period of 29 August, 2001 through 31 December, 2003.
(b) T denotes sample size

0*) denotes 10% signiﬁcance level

0") denotes 5% signiﬁcance level

(...) denotes 1% signiﬁcance level

Itb =71b + 72bdth—1(m)+ 73bat—12 + 31b

where 1,1’ = 1, US,”)0 with 1,” = 0,0therwise
and

2
Its =71“? +yzsdevt_1(m) +73SO't_1 +£ts

where 1,5 =1, UStS)0 and ItS = 0,0therwise

66

2.4.1. Tobit and Censored Least Absolute Deviation Model

Another important kind of limited dependent variable model is one that is continuous
over strictly positive values but is zero for a ﬂaction of the population. Let y be a variable
continuous over strictly positive values but that takes on zero with positive probability.

The Tobit model is most easily deﬁned as a latent variable model where

y. = :30 + xﬂ + u,u Ix ~ Normal(0,0'2) (2.6)

y = maX(0,y') (2.7)

The latent variable y' satisﬁes the classical linear model assumptions where it has a
normal, homoscedastic distribution with a linear conditional mean. Equation (2.7) implies
that the observed variable,y, equals y'when y‘ >0 and 0, otherwise. Maximum
likelihood estimation is used. Both heteroscedasticity and nonnormality result in the
Tobit estimator ,6 being inconsistent for ,6. This inconsistency occurs because the
derived density of y given x hinges crucially on y' | x ~ Normal(xﬂ,0'2). Tests for

heteroscedasticity and non-normality in the latent variable equation are easily

constructed.

A useful test for heteroscedasticity is suggested by Wooldridge (2002) and obtained
by assuming Var(u |x)=0'2 exp(26), where z is a lxq subvector of x (2 does not
include a constant). The q restrictions H0 :6 = 0, can be tested using the Lagrange

Multiplier (LM) statistic. We can also construct tests of non-normality that only require a

standard Tobit estimation. The most convenient of these are derived as the conditional

67

moment tests. See Pagan and Vella (1989). Powell (1984) suggests that a least absolute
deviation estimation is a natural approach for censored data when the assumption of

normality of the errors is suspect. It is possible to estimate ,6 consistently without

assuming a particular distribution for uand without even assuming that uand x are
independent. Consider again the latent variable model , but where the median of u ,given

x, is zero and

y° = xﬂ + u, Med(u | x) = 0 (2.8)

Med( y Ix) = max[0,Med(y‘ lx)] = max(0,xﬂ) (2.9)

The equation (2.8) implies that Med ( y' Ix) = xﬂ , so that the median of y' is linear in x.

Importantly, equation (2.8) holds up under assumption (2.9) and no further distributional

assumptions are needed. Assumption (2.9) suggests estimating ,3 by solving

mﬂin(1/N)Z| y, — max(0, xiﬂ) | (2.10)

i=1

The 1 cast a bsolute d eviations e stimator for the c ensored r egression m odel (CLAD)

minimizes the sum of absolute deviations of y, ﬂom max(0, x,,B) over all ,6.

To obtain the CLAD estimates we use the methodology described in De Jong and

Herrera (2004). Thus, we use an iterative linear programming algorithm, which was

68

proposed by Buchinsky (1994). This procedure amounts to ﬁrst solving the linear

programming (LP) representation of the optimization problem

 

ngin{T1_1;[%sgn(y, -/3'x.)(y. mm} (2.11)

to obtain the 5(1) estimates. Then solve the LP problem for 5(2) using the observations

for which I: (wx, > 0. This procedure is repeated until the set of observation used in two

consecutive iterations is the s arne. Standard e rrors for 3 are 0 btained according to the
method employed by De Jong and Herrera (2004 ). The reported standard errors for the

Tobit estimates are the quasi-maximum likelihood standard errors.
2.4.2Estimation Output

In this section, we ﬁrst estimate the reaction function using a Tobit model and check
for homoscedasticity and normality. The empirical evidence shows that we have a
heteroscedasticity and non-normality problem. Then we estimate the reaction function
using a C LAD m ethod where the n ew e stimator is b oth c onsistent and a syrnptotically
normal. F irst, w e e stimate the T obit m odel for a m anaged ﬂoat p eriod. T he empirical
evidence is in Tables 2.3 and 2.4. We take the same deviation ﬂom the moving average
values as we use in probit model. Deviation ﬂom both the short- and medium-term
appears to have a signiﬁcant effect Granger causing the sale of US dollars. The sale of

US dollars is decreased by a deviation ﬂom the long-term and conditional volatility .

69

While the purchase of US dollar is increased by the deviation ﬂom long term, both the
deviation from short-term and volatility has no effect on the p urchases of U S dollars.
When we ﬁnd that Jarque-Bera and LM tests indicate both non-normality and
heteroscedasticity, we used CLAD method. With respect to the purchase of US dollars,
CLAD results are the same as Tobit results except the volatility appears to have a
signiﬁcant effect Granger causing the purchase of US dollar. With respect to the sale of
US dollars, contrary to Tobit results, deviation ﬂom the short- and long-term appear to
have no signiﬁcant effect. In all these estimations we ﬁnd that a lagged value of

intervention increases intervention next day.

70

ﬂ
Table 2.3: Estimation of Tobit Intervention Model

 

Purchases of US Dollars Sales of US Dollars
Coefﬁcient Standard Error Coefﬁcient Standard Error
c -61.95 13.65 -1 13.56 18.93
DEVMs -8.56 9.55 26.35 ‘ 15.16
DEV,_,25 -35.31 10.91 49.75 15.52
DEV,_,5° 18.74 5.97 -24.59 6.39
6,,2 14.11 24.60 -90.46 " 44.75
1,," 0.66 0.07
It-1' 0.77 0.07 ‘
LM-Test‘b’ 823 926
skewness 4.10 5.03
kurtosis 27.54 48.38
Jarque-Bera 33755 108877
T“) 1209 1209

0" Full period of 15 March, 1995 through 31 December, 1999.
(b) Lagrange Multiplier statistic to test for heteroscedasticity.

(f) T denotes sample size
( )denotes 10% signiﬁcance level

(:2 denotes 5% signiﬁcance level
( )denotes 1% signiﬁcance level

y. = A, +xﬂ+u,u Ix ~ Norma1(0,0'2)

y = maX(0,y')

where yequals y. when y*>0 and 0, otherwise.

71

—a
Table 2.4: Estimation of Powell Censored Least Absolute Devration )

Purchases of US Dollars Sales of US Dollars
Coefﬁcient Standard Error Coefﬁcient Standard Error
0 -13.07 1.61 -102.85 25.86
DEVM‘ 1.13 1.01 -4.87 13.55
015v...25 -7.22 1.21 30.24 " 15.42
new,“ 4.83 0.66 -O.88 8.40
o,,,’ 30.31 2.57 90.22 " 44.22
1,," 0.37 0.01
lot“ 0.51 0.04
M
kurtosis 24.36 16.37
Jarque-Bera 25521 10266
7"” 412 249

 

(a) Full period of 15 March, 1995 through 31 December, 1999
(b) T denotes sample size

(*) denotes 10% signiﬁcance level

('7 denotes 5% signiﬁcance level

(...) denotes 1% signiﬁcance level

y'=x,B+u, Med(u|x)=0

Med (y I x) = max[0, Med ( y. |x)] = max(0,x,6)

mﬂin(1/N)Zl y, — max(0,x.-ﬂ)l

72

Tables 2.5 and 2.6 gives the empirical evidence for both Tobit and CLAD estimations
during the ﬂee ﬂoat regime. First we estimate a Tobit model and ﬁnd that sale of US
dollars is determined by the current volatility of the exchange rate and deviation ﬂom
long-term. With respect to the purchase of US dollar while deviation ﬂom long-tenn
decreases the purchase of US dollar, deviation ﬂom both short- and the medium term
increases. Lagged values of intervention increase the probability of intervention next day.
When we ﬁnd evidence for heteroscedasticity and non-normality, we estimate CLAD
and contrary to Tobit results, we ﬁnd that neither the representative deviations nor
volatility has an effect on the sale the US dollars. The other interesting difference
between the Tobit results is with respect to the purchases of US dollars. Deviation ﬂom
the short-term has no effect on the purchase of US dollar. The sample size in ﬂee ﬂoat
regime is smaller than the one in managed ﬂoat. So we suspect small sample size bias for
CLAD estimatons. Khan and Powell (2001) suggests two-step estimators for these
models to overcome the ﬁnite sample bias. In this respect an interesting extension of this
research would be to use the two-step estimators method suggested by Khan and Powell

(2001) in further research.

73

 

Table 2.5: Estimation of Tobit Intervention Modelm-

 

 

Purchases of US Dollars Sales of US Dollars
Coefﬁcient Standard Error Coefﬁcient Standard Error
C -136.79 40.56 -58.58 17.44
DEVH5 12.20 ' 7.04 1.63 1.26
DEV,_,5° 12.39 4.54 -4.87 1.79
DEVHm -20.68 5.58 4.82 1.82
6,,2 4.49 13.91 5.32 1.69
1,,b 0.16 " 0.07
Itt' 0.69 0.24
LM-Test‘b’ 341 121
skewness 11.98 23.81
kurtosis 181.48 573.53
Jarque-Bera 793204 8017040
T“) 587 587

 

1" Full period of 29 August, 2001 through 31 December, 2003.
(b) Lagrange Multiplier statistic to test for heteroscedasticity.
“’ T denotes sample size

(') denotes 10% signiﬁcance level

0') denotes 5% signiﬁcance level

(...) denotes 1% signiﬁcance level

y. = ﬂo + xﬂ + u,u | x ~ Normal(0,0'2)
y = maX(0.y')

where yequals y‘ when y*>0 and 0, otherwise.

74

 

Table 2.6: Estimation of Powell Censored Least Absolute Deviation“)
—

Purchases of US Dollars Sales of US Dollars
Coefﬁcient Standard En'or Coefﬁcient Standard Error
0 -6.86 13.19 -1044 15.92
01va5 1.18 3.62 0.04 0.69
015v...so 6.92 2.29 -1.08 1.66
new” 6.92 1.91 1.04 1.55
of -1054 3.91 0.68 1.44
1,," 0.06 " 0.02
Iii' 0.90 0.11
kurtosis 139.18 18.25
Jarque-Bera 458938 5946
7"” 186 77

 

“0 Full period of 29 August, 2001 through 31 December, 2003.

(b) T denotes sample size

1') denotes 10% signiﬁcance level
('1') denotes 5% signiﬁcance level

(...) denotes 1% signiﬁcance level

y'=x,B+u, Med(u|x)=0

Med (y Ix) = max[0, Med ( y. I x)] = max(0,x,6)

mpin(l / 07);] y,- — max(0, x716) I

75

2.5 Conclusion

This chapter investigates the motivations for central bank intervention. We estimate
the reaction function for b oth the m anaged float and free float regime separately. The
empirical evidence gives different result for different regimes. First the probit model for
managed ﬂoat was estimated and it was found that the probability of sales of US dollars
is increased by a deviation ﬂom both the short- and medium-term and the probability of
purchases is increased by the deviation ﬂom the long-term. The probability of
intervention is not determined by the conditional volatility for managed ﬂoat and Tobit

estimates give similar results.

Contrary to managed ﬂoat period, both tobit and probit estimations indicate that the
central bank is more responsive to volatility during a ﬂee ﬂoat regime. While the sales of
US dollars is determined by volatility and the deviation ﬂom long-term, purchase is

determined by deviation ﬂom short— and medium term.

When we ﬁnd non-normality and heteroscedasticity in Tobit results, we estimate all
the reaction functions using a CLAD method so that we have consistent and
asymptotically normal estimators. For managed ﬂoat, different ﬂom the Tobit results, we
ﬁnd evidence that purchases of US dollar is increased by volatility and sales are only
determined by deviation ﬂom medium term. For ﬂee-ﬂoat regime, we ﬁnd that neither
the deviation ﬂom the 5,50 and 125 days moving average values nor volatility have an

effect on the sales of the US dollars. In all the estimations, the common ﬁnding is that

76

lagged value of intervention increases the intervention next day for both the managed and

ﬂee ﬂoat regime.

The sample size in ﬂee ﬂoat regime is smaller than the one in managed ﬂoat. So we
suspect small sample size bias for CLAD estimatons. Khan and Powell (2001) suggest
two-step estimators for these models to overcome the ﬁnite sample bias. In this respect an
interesting extension of this research would be to use the two-step estimators method

suggested by Khan and Powell (2001) in further research.

77

CHAPTER 3

Threshold Non-linear Estimation of the Central

Bank Intervention Reaction Function

3.1 Introduction

In chapter two, the intervention reaction function is estimated using probit, tobit and
censored least absolute deviation methods . In this chapter, we adopt a different approach
and use a threshold model, with a view to extending our empirical understanding of the
detenninants of the Central Bank of Turkey‘s intervention. This method is ﬁrst used by
Jun (2004). Jun(2004) mentions that the ﬂiction method adopted by Ahnekinders and
Eij fﬁnger (1996) and Kim and Sheen (2002) is not better than a simple linear model. A
friction model involves specifying three separate distributional assumptions for the
intervention series, and corresponds to the three different states of the intervention
outcome. This approach allows a direct modeling of the relationship between the
interventions and their determinants . The central bank is assumed to react to market
conditions and constraints, but only after an intervention threshold is reached. The

threSholds may differ for positive and negative interventions (purchase/sale of the foreign

78

currency) and these may be estimated. Jun (2004) tests the ﬂiction hypothesis with a
more ﬂexible model using a threshold model that does not restrict the dependent variable
to zero in the middle regime. In contrast to the friction model, the threshold variable that
determines the regimes is one of the measures of the disorderly market conditions and it
can be estimated by the method of least squares without assumptions regarding error
distribution. Residual based test for misspeciﬁcation are also applicable using this
modeling approach. Section 3.2 gives a detailed description of the speciﬁcation,
' estimation and test strategies of a threshold model in the context of intervention (Hansen,
1999, 2000). The empirical results with respect to the threshold model are given in

section 3.3 and section 3.4 gives a brief conclusion.

3. 2 Estimation Of Threshold Models

The linear model of central bank reaction function is
y: =Xtﬂ+£t (3.1)
where y, d enotes the actual p urchases (sales) 0 f U S dollar and an m -regime threshold
model allows the parameter vector ,8 in the linear model (3.1) to change m times based
on the values of the threshold variable 77,. A three regime threshold model can be written

as

yt = xtﬂl-K’k 5 7t) + xtﬂz-IU’I S 771 S 72) + xtﬂ3-“ﬂt Z 72) + 81 (32)

where I (.) is the indicator function, and ,6,- = (ﬁt-0,6,1 ...... ,Bl-k) , i = 1,2,3.

79

If x, consists of lags of y, only, it is called a threshold autoregressive (TAR) model. In
addition, if 77, is one of the lagged y,, the model becomes a self- exciting threshold

autoregressive (SETAR) mode]. Some exogenous variables are used, such as deviation
and volatility measures as explanatory variables, in addition to the autoregressive terms

of y,. The model is called a threshold model (Hansen, 2000). In our analysis, the

threshold variable 77, is one of the exogenous variables. Speciﬁcally, we compare the

dev(,_,) and 004,2 , and choose the one that minimizes the sum of squared residuals as
the threshold variable. The parameters (81,82,63,7,,72) can be estimated using

conditional least squares. If the thresholds are given, the model becomes linear in the rest

of the parameters and can be written as

y, = 3,6 + 8', (3.3)
where (9 = (,61, ,62, ,63) and the 3(k+1) row vector .9, is
37 =(x713-1(777 S7’1) 1115 1(71 5 '77 S7’2) xtﬂ-“Ut 2 72)) (3-4)

The 6 in (3.3) can then be estimated by OLS. In deﬁning the set of threshold
observations

Q = {77, It =1,....T} (3.5)
For each pair of (77,.,77j)e(22 where i,j=1 ...... T and 77,071., substitute (77,,77j) for

(71 , 72) in (3.4). Estimating (3.3) by using OLS, the sum of squared residuals S(77,., 77 j) is

obtained and is deﬁned as
T 2
8071.77,) = t21(yt - 3149) (3.6)

80

Then the pair of (77,, 77 ,) that minimizes the sum of squared residuals will be the threshold

A A A A A

estimates (7,, 7,). The estimates of the other parameters (,6, , ,6, , ,6 3) , are obtained as the

OLS estimates of equation (3.2) with 7, = 7, and 7, = 7,. This procedure requires
T (T — l)/ 2 OLS regressions. Since we have two candidates for the threshold variable,
the whole estimation process requires T (T — l) OLS regressions. Although this is not a

problem in the estimation stage, it becomes a critical problem in the hypothesis testing
where the computation o f b ootstrap p -values requires thousands o f replications o f this
procedure. In order to reduce this computational burden, Hansen (1999) proposes a two-

step estimation procedure where the thresholds are estimated sequentially, one by one.

First we estimate a two-regime model as

yt = Mal-[(777 5 7) + 35702-1077 2 7’) + 9': (3-7)

A A A

The threshold estimate 7 for the two regime models is the same as either 71 or 7, and

useful in reducing the running time for the estimation of the three regime model of (3.2).

A

' It is also true that a, = ,6, if 7 = 7,,or a, = ,63 if 7 = 7,. The two step approach proposed

by Hansen (1999) consists of the following steps :

A

1) Estimate the two regime model and obtain 7 which is the element of
Q = {77, It = 1,....T} that minimizes the sum of squared residuals.

2) Estimate the three regime model (3.3) after setting the 7,and7, in (3.4) as either

(7. 77.) or (2.7) for each 77. 6 0. By assumvtion. 7. (72.

81

Hansen (1999) mentions that if the true model is a three-regime model but a two-

A

regime m odel is e stimated, 7, will b e c onsistent for o ne p air 0 f the (7, , 7,). Then if

(7, , 7,) is estimated by least-squares enforcing that one element of 7 equals 7, , then the

A

second stage estimate 7, will be consistent for the remaining element of the pair (7, , 7,).

Furthermore, this method is iterated at least once; that is, (7,, 7,) is estimated by least

A

squares enforcing the constraint that one element of 7equals 7, , yielding a reﬁned

A

estimate 7, . This iteration makes the estimates of (7,, 7,) obtained by the two-step

method asymptotically efﬁcient.

If there are T elements in Q for each of the N candidates of the threshold variable,
then the two-step approach requires (2T-1) N regressions. In order to estimate the multi-
regime model of (3.2), some restrictions must be imposed on the range of 7,and7, so
that each regime has at least the minimum required number of observations . One obvious
requirement is that the number of observation of each regime, 7:. for i = 1,2,3 must be
greater or equal to the number of explanatory variables, k+1. Another more important
requirement is concerned with the asymptotic properties of the estimators and test
statistics. For a linear model, we can rely on consistency of the OLS estimator if T is

large. Similarly, with three regime m odel o f (3.2), w e n eed to h ave large I, for each

regime because the conditional least squares procedure is equivalent to splitting the

sample into three sub-samples and then estimating ,6, using only T, observations. This

82

involves three smaller regressions and is faster than running one large regression with

equation (3.3). Hansen (1999) explains that it is necessary to have T, / T 2 r for some
r)0 as T —> 00. In practice, however, it is inevitable to choose 2' somewhat arbitrarily.

Hansen (1999) suggests r = 0.10.

The next step is to test if the nonlinear model in (3.2) is a better speciﬁcation than the

linear model in equation (3.1). This is equivalent to testing the null hypothesis of
H01=ﬂ1=ﬂz =ﬂ3=ﬂ (3-8)

Hansen(l999) suggests a test statistics of

S _
F = T(—LT:—T—) (3.9)

where S , is the sum of squared residuals ﬂom LS estimation of the linear model (3.1)

and S, is the sum of squared residuals ﬂom the three-regime model (3.2). When the

thresholds are known, F is asymptotically equivalent to the usual F statistic. Since the
thresholds are unknown and not identiﬁed under (3.8), however, F follows an unknown
asymptotic distribution. Bootstrapping methods are relied on to compute the p-values

with and without the conditional heteroscedasticity assumption.

Under the h omoscedastic e rror a ssumption, a s et 0 f b ootstrap e rrors a re 0 btained,

with E = {5, It: 1, ..... T} by randomly drawing T times with the replacement ﬂom the
OLS residuals é={é, |t=1, ..... T} of the linear model (3.1). A set of data on the

dependent variable is generated then by

83

A

~ ~ ~

yr = X: 6+ 3t (3.10)

~ ~ ~ ~

where x, = (1,dev(,_1)m,0(,_1)2,y,_1,y,_2,....,y,_p )' and 6 is the OLS estimate of

equation (3.1). Substituting )7, for y, in the linear model and the threshold models, the

models are re-estimated to get one value of E which is

f =T(§°—§—§‘—) (3.11)

where S, and S, are the sum of squared residuals ﬂom the linear model and the threshold

model, respectively, with bootstrap data. Out of 1000 replications, the proportion of F‘

greater than F is the approximate p-value.

Under heteroscedastic error assumption, the procedure is a bit more complicated

because we have to impose heteroscedasticity on the bootstrap errors a. First each

A

element of the OLS residual vector 5 is divided by an estimate of the conditional standard

deviation (Jiz) to obtain a set of homoscedastic errors of

~ ~ ~

8={£t |8,=s,/ h,,t=l, ....... T} (3.12)

The conditional variance estimate, hr , is obtained as the ﬁtted value ﬂom an auxiliary

A

regression of 8,2 on x,2 = (l,[deV(,_1)(m)]2,[a(,_1)2]2,y(,_1)2,y(,_2)2, ...... ,y(,_p)2)'

84

A

ef=x35+m (3B)

A A

h=x35 0J0

A

where U, is an error term and 6 is the vector of OLS estimates in that auxiliary

regression. Now the random draws are ﬂom these standardized errors of 8 and the t-th

bootstrap error 6‘: is

~ z ~

8t =8! ht (3.15)

~A

~

WhCI‘C ht =Xt26 and xt =(l,deV(,_1)2,0’(,_1)2,y,_1,y,_2, ...... ’yt—p). once the

value of 6, is given, the value of the dependent variable y, is computed by (3.10). It is

A

important to note that h, :1: h, and x. at x,. x, contains lags of y, , h, , a, as well as y, ,

and must be computed recursively. The rest of the bootstrap procedure is the same as the

homoscedastic case.

If the linearity hypothesis is rejected in favor of the threshold model, Hansen (1997,
2000), in turn, provides a method to construct an asymptotically valid conﬁdence interval
for the threshold of the two regime model. Unfortunately, to date the asymptotic
distributions of the threshold estimators ('71, 72) are not developed for three or higher

order models.

85

3.3 Tests for Nonlinearity and Estimation Output of Reaction Function

We estimate the reaction functions for the Central Bank of Turkey for both managed
ﬂoat and ﬂee ﬂoat regime. For each of the subsamples, the two regime and three-regime
threshold reaction functions are estimated using GAUSS programs. This estimation
precedes hypothesis tests, but begins with the results of the tests whether two-regime or

three regime threshold nonlinearity exists in the reaction function.

The tests of nonlinearity for a managed ﬂoat regime are presented in Table 3.1. The
test statistics F 12 for a o ne-regime, v ersus a two-regime linear model is 2 7.35. The p -
values are computed as the proportion of those bootstrap simulations out of 1000
replications that have the F -statistic larger than 27.35. When the errors in the linear model
are assumed to be homoscedastic, the p-value is 0.01. Therefore,we reject (fail to accept )
the null hypothesis of linearity against a two-regime threshold-nonlinearity at 5% level.
Correcting for heteroscedasticity, the p-value is about 0.37 and we do not reject the null
of linear reaction function. Against the three-regime alternative, the F13 statistic ﬂom the
estimation is 50.09. The bootstrap p-value is 0.02 with homoscedasticity. Therefore we
reject the null of the linear reaction function at 5 %. With heteroscedasticity, the p-value
is 0.27 and fail to reject the linearity. Since we cannot reject linearity against three-
regime nonlinearity, it is not suprising to ﬁnd that the three-regime model is no better
than a two-regime model. The F23 statistic is clearly signiﬁcant with homoscedasticity

assumption but not with the heteroscedasticity assumption for the regression errors.

86

Table:3.1

Test for Non-linearity

Managed Float Period (February 7, 1995- December,31 1999)

F-statistlcs

p-values

 

Homoscedastic Heteroscedastic

27.35 0.01 0.37
50.09 0.02 0.27
22.24 0.06 0.2

87

The estimation results can be found in Table 3.2. The second column is for the linear
model, the next two columns are for the two-regime threshold model, and the last three
columns contain the estimation results with the three-regime model. As for the linear

model, among the three measures of the disorderly market conditions (devHS, devH25 ,
0',_,2 ), the estimate of devH5 is signiﬁcant at 5 % and the estimate of 0',_,2 is signiﬁcant

at 10%. The coefﬁcients of devH5 and devH25 have an expected negative signs implying

that the central bank‘s intervention tends to be against the wind. On the contrary, the
coefﬁcient of 0',_,2 is positive. Out of the two lags of the dependent variable, both of

them are signiﬁcant at 1% level. Overall, the implication is that recent interventions
increase the expected amount of intervention in the near future. These two lags are
enough to eliminate the serial correlation in the residuals. Ljung-Box test statistics
indicate that there are no serial correlations in the residuals up to order of 10 (Q00) = 13.2,

with a 5% critical value of 18.31).

However, there is strong evidence for heteroscedasticity (Q2(10)=107.48). A separate
regression of the squared residuals on squared regressors (as reported in table 3.3),
indicates that the errors are heteroscedastic and, therefore, White‘s heteroscedasticity

consistent standard errors are reported in table 3.2.

88

 

Table 3.2: Estimation of Threshold Intervention Reaction Function'I (n,=dev,,,s)

Variable

Constant

Dev,_,(5)

2
91-1

Yi-l

yt-Z

T
R2
F,.=0
71
72
2
R (.11)
Skewness
Kurtosis

010
0’...

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Linear Two-regime - Three-regime
Regime I Regimejz Regime I Regime 2 lRegime 3
0.04 0.03 -0.05 -0.01 0.72 ' -0.05
(0.07) (0.09) (0.25) (0.10) (0.43) (0.25)
-0.13 ” -0.07 -0.12 -0.104 -0.83 ' -0.12
(0.06) (0.09) (0.15) (0.13) (0.43) (0.15)
-0.01 -0.002 0.002 0.015 -0.04 0.002
(0.02) (0.03) (0.05) (0.03) (0.07) (0.05)
0.26 ' 0.19 0.57 " 0.192 0.34 0.57 "
(0.15) (0.27) (0.27) (0.31) (0.91) (0.27)
0.37 0.29 0.59 0.344 0.26 0.59
(0.04) (0.04) (0.10) (0.05) (0.06) (0.10)
0.18 0.18 0.13 0.291 0.09 ' 0.13
(0.04) (0.04) (0.08) (0.06) (0.05) (0.08)
1234 939 295 545 394 295
I 0.25 0.19 0.37 0.26 0.14 0.37
84.94
0.94 0.57
0.94
0.25 0.27 0.28
-0.58 -0.57 -0.62
13.31 12.6 12.34
13.2 11.89 13.63
107.48 122.19 114.88

 

 

 

(a) Full period of 7 February, 1995 through 31 December, 1999.
a” Numbers in parathesis are heteroscedasticity adjusted standard errors.

“’ F...=0 tests overall signiﬁcance except the constant.

(:1 denotes 10% signiﬁcance level
(“i denotes 5% signiﬁcance level
( ’ denotes 1% signiﬁcance level

yt = 101511077 5 71) + xtﬂz-IO’I S 777 S 72) + 367163-1077 ->- 72) + 81

where I (.) is the indicator function, and ,6,- = (6,0,6,1 ...... 6,7,) , i =1,2,3.

89

Table 3.3. Least Square Estimation of Conditional Variance

Managed Float Period (Februag 7, 1995- December,3l 1999)

 

 

 

Variable Estimate Standard Error
Constant 0.87 0.15
Dev(5),.,2 0.08 0.05
Dev(25),_,z -0.012 0.01

Vol...2 .052 0.22

y..,2 0.11 0.06

yez’ 0.04 0.03

F-Test 40.19

(p-value) 0.00

90

Tests for nonlinearity (Table 3.1) show that a rejection of both two-regime and three-
regime models versus the linear model. Despite this result, it is worth mentioning that

with two regime model, the optimal threshold variable is estimated to be dev,.15 and the
point estimate of the threshold 77, is 0.94. This positive threshold indicates that the
strongest non-linearity in the central bank‘s reaction function exists when the dollar
appreciates rapidly. This indicates that, in light of the history of the exchange rate policy
under managed ﬂoat, the central bank is more sensitive to appreciation of the dollar. In

the case of the three-regime model, the estimate of the second threshold is also positive.

The test results for both the non-linearity and the estimation output of the reaction

ﬁrnction for the ﬂee ﬂoat regime are given in tables 3.4 and 3.5.

91

 

Table:3.4 Test for Non-linearity

Free Float Period (April, 10 2001- December,3l 2003)

 

F-statistics p—values
Homoscedastic Heteroscedastic
F12 25.64 0.05 0.37
F” 48.29 0.39 0.41
F23 21.83 0.33 0.25

92

The test s tatistics F 12 for a o ne-regime linear m odel v ersus a two-regime model is
25.64. The p-values are computed as the proportion of those bootstrap simulations out of
1000 replications that have the F-statistic > 25.64. When the errors in the linear model are
assumed to be homoscedastic, the p-value is 0.05. Therefore, we reject (fail to accept )
the null hypothesis of linearity against a two-regime threshold-nonlinearity at 5% level.
With correction for heteroscedasticity, the p-value is about 0.37 and we do not reject the
null of linear reaction function. Against a three-regime alternative, the F13 statistic from
the e stimation is 4 8.29. The b ootstrap p -value is 0 .39 w ith h omoscedasticity and 0 .41
with heteroscedasticity. and we fail to reject the linearity. Since we cannot reject linearity
against three-regime nonlinearity , it is not suprising to see that the three-regime model is
no better than a two-regime model. The F23 statistic is clearly insigniﬁcant with or

without the heteroscedasticity assumption for the regression errors.

The estimation results are given in table 3.5. As for the linear model, among the three

5

measures of disordely market conditions (dev,.15 , devH2 , 0’,_,2) , only the estimate of

0',_,2 is signiﬁcant (1 % level) and has an expected negative sign implying that the

central bank‘s intervention tends to be against the volatility, but not the deviation ﬂom
the trend. Out of the three lags of the dependent variables, all of them are signiﬁcant at
1% and 5%. Overall, the implication is that recent interventions increase the expected
amount of intervention in the near future. These three lags are enough to eliminate serial
correlation in the residuals. Ljung-Box test statistics indicate that there are no serial
correlations in the residuals up to order of 10 (Q(10) = 13.35, 5% critical value 18.31).

Also, there is no evidence for heteroscedasticity (Q2(10)=0.12). A separate regression of

93

the squared residuals on squared regressors, as reported in table 3.6 also indicates that the
errors are not heteroscedastic. As a result, a two-regime threshold is preferred to linearity

and three regimes.

94

 

Table 3.5: Estimation of Threshold Intervention Reaction Function' (1156,42)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Variable Linear Two-regime Three-regime
I Regime I Regime T Regime 1 Regime 2 Regime 3

Constant 0.13 0.29 0.08 0.31 -0.002 0.015
(0.03) (0.40) (0.03) (0.40) (0.36) (0.03)

Devm‘” -0.03 -004 -002 -004 0.08 “ 003 °
(0.02) (0.09) (0.02) (0.09) (0.04) (0.02)

Dev...” -0002 -0.1 ' 0.002 -010 ‘ -002 0.011
(0.01) (0.06) (0.01) (0.06) (0.01) (0.01)

6.12 -0.05 ‘” -O.6 -004 -0.65 0.14 -003
(0.01) (1.05) (0.01) (1.04) (0.53) (0.01)

y... 0.13 0.04 " 0.32 0.04 “ 1.02 0.21 "
(0.05) (0.02) (0.12) (0.02) (0.34) (0.09)

y.2 0.08 0.24 0.04' 0.24 -0012 0.03
(0.03) (0.20) (0.02) (0.20) (0.01) (0.07)

y.., 0.1 “ 0.24 0.06 " 0.24 -0.003 0.22 “
(0.04) (0.22) (0.03) (0.21) (0.01) (0.09)

r 687 161 526 162 200 325

R2 0.11 0.05 0.2 0.05 | 0.14 | 0.34

F..=o 14.06 |

7. 0.45 7 0.45

72 0.89

R,,.,, 0.11 0.14 0.16

Skewness 10.72 10.61 10.81

Kurtosis 159.07 157.21 162.93

0,, 13.35 8.98 4.62

0‘... 0.12 0.13 0.14

 

 

 

1" Full period of 10 April, 2001 through 31 December, 2003.
m Numbers in parathesis are heteroscedasticity adjusted standard errors
“’ Fw=0 tests overall signiﬁcance except the constant

"1 denotes 10% signiﬁcance level
1' ’ denotes 5% signiﬁcance level
(..., denotes 1% signiﬁcance level

yt = xtﬂl-K'h 5 77) + 10.321015 777 S 72) + 157163-1071?- 72) + 87
where I(.) is the indicator function, and 6, =(6,-06,-1 ...... 6,7,) , i =1,2,3.

95

 

Table 3.6 Least Square Estimation of Conditional Variance

Free Float Period (April, 10 2001- December,31 2003)

 

Variable Estimate Standard Error
Constant 0.760 0.430
Dev(5)..,2 -0003 0.020
Dev(25)...2 -0001 0.002

Vol.4: .0001 0.001

y...2 0.010 0.002

m’ 0.001 0.002

y,_32 0.003 0.002

F-Test 0.11

(p-value) 0.99

 

 

96

With the two-regime model, the optimal threshold variable is estimated to be
0',_,2 and the estimate of the threshold variable is 0.45. This threshold indicates that the

strongest non-linearity exists in the central bank‘s reaction function when there is

volatility. This supports the claim that the central bank is more sensitive to volatility in a
ﬂee ﬂoat regime. Most of the observations are in regime two. The coefﬁcient of a,_,2 is

signiﬁcant at a 1% level in this regime. This indicates that central bank‘s intervention is
against the volatility when the volatility is greater than the threshold variable. This

indicates that the central bank is more responsive to high levels of volatility in a ﬂee ﬂoat
regime. Similar to the two regime model, the coefﬁcent of 0',_,2 is signiﬁcant at a 1%

level in regime three. This result supports the ﬁnding in two-regime model such that

central bank is more responsive to high levels of volatility.

97

3.4 Conclusion

The empirical evidence ﬂom the threshold model gives consistent results with the
policies implemented by the central bank during the period investigated. For a managed
ﬂoat period, we reject both the two regime and three regime model. The linear models
give that deviation ﬂom the 5 days moving average and recent interventions matter but
not the volatility. For a free ﬂoat period, under the assumption of homoscedastic errors,
we fail to accept linearity against the two regime model and both volatility and recent
interventions matter. These ﬁndings are consistent with the announcement of the CBRT
such that the aim is not to affect the level of the exchange rate but rather the excess

volatility.

98

CHAPTER 4

Risk Premium and Central Bank Intervention

4.1 Introduction

There has been an enormous amount of literature concerning the effects of
intervention on currency markets and the motivations for such interventions. These
results are mixed and depend on which exchange rate is analyzed , what sample period is
studied and also what intervention strategy is followed. There is empirical evidence that
indicates that some types of intervention can affect the risk premium in forward markets.

However, the risk premium is not necessarily the intended target of the intervention .

The Forward exchange rate is a contractual exchange rate established at the time of a
transaction that will take place at the maturity time t+l and usually is regarded as the
unbiased predictor of the future spot exchange rate. Contrary to popular theory, empirical
evidence shows that the forward rate is a biased predictor of the future spot rate and/or is
evidence of a risk premium. See Hansen and Hodrick (1980), Hakkio (1981), Baillie et
al.(1983), and Baillie (1989). The forward premium anomaly, where the currency of the
country with the higher rate of interest is more likely to appreciate than depreciate, is

generally r egarded a s o ne 0 f t he m ost important u nresolved p aradoxes in international

99

ﬁnance. Numerous explanations have been proposed to explain the forward premium
anomaly, but, to date, no one has been able to fully explain the available empirical
evidence. See Evans and Lewis (1995), Karninsky (1993), Lewis(1988) Frankel and

Froot (1987), Lewis (1989), Elliot and Ito (1995).

In a study by Bailie and Osterberg (1997a), Hodrick’s model (Hodrick, 1989) is
extended to allow central bank intervention to have a direct effect on the risk premium.
Baillie and Osterberg (1997a) ﬁnd that purchases of US dollars by the Federal Reserve
Bank appear to signiﬁcantly increase the excess dollar denominated returns for both the
DM-S and the Yen-S markets. Consistent with this study, Baillie and Osterberg (2000)
found that the intervention variables affect the risk premium in an analysis where the
relationship between daily deviations ﬂom uncovered interest rate parity and intervention

are investigated by using daily overnight euro-currency deposit rates.

Central Banks use intervention as a policy instrument. Despite its ﬂequent use,
intervention continues to be debated as a policy tool due to the controversy over whether
it can achieve the policy goals of either changing the level of nominal exchange rates or
reducing its volatility. In studies investigating the impact of intervention directly on the
levels of exchange rates, it has generally been found that either intervention has no
signiﬁcant effect or that its outcomes are the opposite of those intended. In the ﬁrst
chapter of this study, the effect of intervention on spot exchange markets is discussed in
detail as well as an empirical analysis of the Turkish spot market. This chapter aims to

investigate the effect of intervention on risk premium and to asses whether intervention

100

helps to explain the forward premium anomaly, as found by Baille and Osterberg (1997,
2000). The analysis is done for Turkish economy, where the economy is small and has
high inﬂation. Section 4.2 describes the details of the model, Section 4.3 gives the data,

Section 4.4 presents the estimation output and Section 4.5 discusses the results.
4.2 Details of the model: Risk Premium and Intervention

The Covered Interest Rate Parity Condition tells us the relationship between spot

rates, forward rates and interest rates.

05,: —st+1> =0, -.1, 1 (44)

s, and f,, corresponds to logarithmic values of spot and forward exchange rates,

respectively. Also i,, denotes the domestic currency return on an l-period risk ﬂee dollar

bond, denominated in terms of domestic currency where as i 3,1 is the foreign currency
return on a risk ﬂee bond denominated in terms of the foreign currency. It implies that
the country with the higher rate of interest has experienced an expected depreciation of
currency. Sometimes, the relationship between forward rates and future spot rates is
simply expressed in terms of the forward rate as being an unbiased predictor of the future

spot exchange rate and is given by

ft,l:EtSt+[ (4.2)

101

where s,,, is the logarithm of the spot exchange rate and f,, is the logarithm of the

forward rate for maturity in time t+l. This is widely rejected by the empirical studies. See
Hansen and Hodrick (1980), Hakkio (1981), Baillie et a1. (1983), Baillie (1989). This has

led to a type of model as in

ft,I = EtSt+1 + 72: (4.3)

where p, is a time dependent risk premium. The dependent variable in this study is the

difference between expected rate of appreciation of the US dollar and the forward

premium or in other words risk premium, which is deﬁned as (Sm, - 8). Note that

*
(St+k — St) — (it ‘it )= St+k — St - (ft ‘57) = St+k ‘ ft (4-4)
Hence,
(4+1. -n>=pt +4.7. <45)

where um, is the rational expectations error associated with using the forward rate to
predict the spot rate k periods and u, is serially uncorrelated for lags greater than k, so
E(utu,+h) = 0 for h > k. This restriction is consistent with u,, following a moving average

process of order k-l .

Baillie and Osterberg (1997a) extend Hodrick’s 1989 model based on a consumption

based asset pricing model, where risk premium depends on the conditional variance of

102

production, money growth rates, consumption’s share of production and intervention

variables.

-aa 2—a 24+ 2—a0' 2*+ 2— 2*+ar —ar*(46)
pt' 1 yt 20y! “300! 4 Q: “5%, “60;, 7ft 81 °

where 0' 2 and 09,2 are the conditional variances of logarithms of production and the

yt
gross grth rate of the currency, respectively. The variable 0'92 denotes the

conditional variance of the share of the currency used for intervention. The intervention
variable 7, = M ' /M , which is deﬁned as the share of currency held by a foreign

govermnent for intervention Operations. Astericks denote foreign country equivalents.
The difference between this and Hodrick’s model is the addition of the conditional means
and variances of the two intervention variables in the risk premium. The model does not
impose any restrictions on w hether o r 11 ot sterilization o ccurs. The m odel is e stimated
ﬂom daily data in order to determine the relatively short-lived effect of intervention on
risk premium. Hence it is not possible to include the variances of production, money
growth rates, and foreign currency holdings as a proportion of money stock. The spot
exchange rate, the forward exchange and the intervention variables, which are all
observed daily , are the variables included in the estimated model. Hence the risk

premium p, in (4.6) is considered to be determined by

(SH-k -ft) = 8t '1' . z 61.81“]. + b0 '1’ blUStb + szStS (4.7)
7:1,21

5t = 2,0', (4.8)

0,2 =co+a6',_12 +6a,_12 (4.9)

103

The ﬁrst two terms on the right hand side of Eq (4.7) corresponds to u, and 8, is a

serially uncorrelated white noise process, and 0,- are the moving average parameters. The

explanatory variables, US,b,US,s include the intervention variables. Conditional

variance in equation (4.10) is represented by a linear GARCH (1 ,1) process.

Bollerslev (1986) introduced the GARCH(Generalized Autoregressive Conditional

2

Heteroscedasticity ) process, which extends the ARCH model to make 0', a function of

lagged values of 0,2 as well as the lagged values of 8,2 . Bollerslev (1986) required all

the coefﬁcients to be positive to ensure that the conditional variance is never negative.

The quasi-maximum likelihood estimation is used.

4.3 Data

In this study, the data is provided by the Central Bank of Turkey, the Istanbul Stock
Exchange and Federal Reserve Bank Board of Governors. The November 1993 through
December 2002 data sample consists of daily spot offer rates, interbank ovemight interest
rates, treasury bill rates, 30-day euro dollar rates and daily intervention variables.
Intervention values are millions of US dollars. This study uses the daily amount of net
dollar purchases (sales), daily spot offer rates and interest rates. The analysis separately
covers both the managed ﬂoat and ﬂee ﬂoat period in terms of exchange rate regime.
The development of a ﬁrtures market is very new to the Turkish Economy. The forward

exchange rate is calculated. The implied forward rate12 is

 

' 360-day is assumed as the basis for interest quoations instead of 365, see Grabbe (1996).
2 The Implied Forward Rate is calculated as given in Grabbe (1996).

104

30

36000
30
))

36000

 

S,(1+i'r( ))

F130 - (4.11)

 

(1+i,(
where F,30 is the daily 30-day forward rate, S, is the daily spot rate as H USD , see

Figure 4.1 and Figure 4.2. i '7 is a proxy the 30 day treasury-bill interest rates for Turkey.
Daily interest rates for treasury bills traded in the secondary market is obtained ﬂom the
Istanbul Stock Exchange. The interest rate of which the treasury bill has the closest

maturity to 30 days is chosen for each day. i, is a proxy for 30-day euro dollar rates.

105

 

j-l-l 1.11

mam—=6 AL u mega—é Hun—55mm wane Hﬁam—v 5 H

 

Ncccccc A?
Hmccccc.%lliiilllitilt
Hcccccc lilililliiiliill.

 

 

 

_ Heeceec
uneeece 11:

 

 

_ Heeeecc 1111111-

 

mccccc

 

 

 

 

 

 

cccccc

 

 

 

 

accccc

 

 

Mccccc

 

 

 

 

l .
W v k ﬂ._s._1_7-__l__ hi ~...in__.__i~_

 

m”

a 6
WWW

 

 

m 6
mm

 

AJ Cam—w. adieu—... was en dram—c mun—5:5“ Em 3:312— en — 2365—3... 58 9.6:”.— E coco—=56... was.

106

 

 

7 mam—=6 Ab u maven Hue—.350 Wane HEW—vac

Hmccccc

 

 

Hmeeeee
Heeeeee -
_ -eeeee

 

 

 

Hcccccc 111:

 

 

 

 

 

 

 

mccccc L1 .1 , 11
cccccc 11|1, - \t\SlL
occccc 1111-
Nccccc

cLlllni

6. a e a 6. m m m

a m
M m m m -W ”M m mm

 

 

 

 

 

 

107

 

 

AJ Cam—w. mg» ”a? 3. cutaway man—5:5“ =3 3: 6612— 3. w 232.52.. 58 93:“: u— coon—.52.. was.

In this study, the forward rate quotations are matched with the future spot rate so that
both represent contracts that would be delivered on the same day. The details of
settlement procedures in the spot and forward markets are discussed in detail by Rich]
and Rodriquez (1977). The important aspect here is the number of working days in the
contract period varies. One reason is that delivery delays often occur around the ﬁrst of
the month. Contracts also are not settled on weekends or on holidays in either of the two
countries for a given exchange rate . This exact matching reveals that for the data used in
this study, k, the number of working days ﬂom the day of the forward quote to the time
of settlement in the spot market varies ﬂom 20 to 26. Since the most common value of k
in our sample is 22, u,, the forecast error is estimated as an MA(21) process. This
analysis has been done for two sub-periods due to the difference in economic policies.
The ﬁrst sub-period covers between August 1, 1994 and November 30, 1999. and the
second one between February 22, 2001 and December 31, 2002. The forecast error for

two sub-periods are shown in Figure 4.3 and in Figure 4.4

108

 

 

j lac—.6 Pu" magma 33 40.6086» 62.018

 

 

 

 

 

 

m Pom 1i- -- 111- -- _ -w -. i --1,

 

 

 

 

 

 

 

 

 

 

 

 

 

 

1c.N_21- - i --..
_WM mxmmmmmmmmmmm

 

 

 

 

memmmMmmMWMMMQ.
AJ Cam—w. mega...— 2—8 3393 2.3.. on dram—v man—5E5 c.» 3: 1612— 3. — cram—5.. 5?. 93am—— uc 236532.. See.

109

 

 

 

U PNM
. 98
w ...—
L Pam

 

 

 

 

..cbm
ue. —
u? H m
LVN
IPNM

 

 

_

 

 

 

 

 

 

_

e M Z W I M M W M

I Feb-01
r-01

[it

 

AJ Cum—w. want»:— 38 3.62—ma 2.3.. a. dram—v nan—=35“ :5 3: 3.32. 3. 3 £335. 33 935: H: USoEra—wuccu.

110

4.4 Estimation Output

The details of the estimated model from the daily risk —premium and intervention
data are given in Table 4.1 and Table 4.2. The model possess estimated moving average
coefﬁcients that approximately decline linearly with the lag. Diagnostic testing of the
model fails to provide evidence for a higher order moving average process. Also a linear
GARCH (1,1) process is found to be an adequate representation of the conditional second
moments for the managed ﬂoat period and a linear integrated GARCH (1,1) is adequate

for free ﬂoat period.

The most interesting aspects of the estimated models in Tables 4.1 and 4.2 concern
the coefﬁcients of the variables associated with intervention. In particular, unlike Baillie
and Osterberg (1997a) , both purchases and sales of US dollars by the Central Bank of
Turkey appear to have no effect on the size of risk premium for TL/USD for the free ﬂoat
period. Similar results are found for the managed ﬂoat period but the buying of US
dollars appear to have signiﬁcant effect at a 20 percent signiﬁcance level. This ﬁnding is
expected to b e the r esult o f h igh inﬂation in T urkish E conomy. E fforts of d isinﬂation
were not successful through 1990s and stability in the foreign exchange market was
uncommon. Under these circumstances, factors affecting the interest rates are related to
stability in both the domestic market and government debt management. The Central
Bank of Turkey aimed at achieving stability in the markets. Under these circumstances,

no relation is expected between risk premium and intervention.

111

Table 4.1: Estimation of Intervention/Risk Premium Model: TL/S “0

Conditional Mean Parameters

 

 

 

 

Coefﬁcient Standard Error Coefﬁcient Standard Error
Do 0-0 0.01 0-0 0.004
mus") 0.0 0.0002
bz(US‘) 0,0 0.0001
91 0.9 0.04 0.9 0.055
02 0.7 0.07 0.7 0.062
03 0-7 0.12 0.7 0.098
04 0.7 0.14 0.7 0.121
95 0-7 0.13 0.7 0.11
06 0.7 0.11 0.7 0.108
07 0.6 0.08 0.6 0.078
0. 0.5 0.06 0.5 0.065
09 0.6 0.05 0.6 0.059
9m 0.6 0.05 0.6 0.058
911 0.5 0.07 0.5 “ 0.075
912 0.5 0.06 0.5 ” 0.065
013 0.4 0.06 0.4 0.063
914 0.5 0.06 0.5 0.064
015 0.4 0.06 0.4 0.059
0m 0.3 0.06 0.3 ‘ 0.064
0n 0.2 0.05 0.2 0.058
01; 0.2 0.05 0.2 0.058
919 0.2 0.05 0.2 0.056
920 0.2 0.05 0.2 0.049
”21 0-1 0.0;; 0-1 " 0.042
Conditional Variance Parameters
to 0.0 0.0 0.0 ' 0.0
a 0.3 0.1 0.3 " 0.1
p 0.6 " 0.1 0.6 0.1
Skewness l .07 1 .00
Kurtosis 23.34 22.45
Q20 24.20 22.44
oz; 13.36 13.23
T 1347 1347

 

m Full period of 1 August, 1994 through 30 November, 1999.
1') denotes 10% significance level
(..) denotes 5% signiﬁcance level
(..., denotes 1% signiﬁcance level

S
(s... -f.) = a, + z 19 jg,_ j + b0 +b1US,b + bZUS,
j=l,21

2 2 2
a, =w+aa,_1 +ﬂa,_1
is iid (0,1)

8t =Zt0't With zt

112

Table 4.2: Estimation of Intervention/Risk Premium Model: TL/S ‘"

Conditional Mean Parameters

 

 

 

 

 

Coefficient Standard Erron Coefficient Standard Error
bo ~0-O35 " 0.016 -0.03 '° 0.016
b,(US") 0.00 0.0007
b2(US') 0.00 0.0014
Or 0.99 0.058 0.99 0.060
02 0.80 0.067 . 0.80 0.068
03 0.83 0.081 0.85 0.081
04 0-70 0.083 0.70 0.082
05 0.75 0.085 0.76 0.088
06 0.74 0.092 0.75 0.097
07 0.70 0.096 0.70 0.100
0a 0.65 0.096 0.66 0.099
09 0.75 0.086 0.75 0.092
9m 0.87 0.081 0.87 0.079
0n 0.76 0.067 0.77 0.069
912 0.76 0.076 0.77 0.082
013 0.79 0.087 0.80 0.085
014 0.68 0.063 0.67 0.063
915 0.66 0.078 0.68 0.083
9.6 0.61 0.082 0.60 0.081
0n 0.56 0.090 0.55 0.086
01. 0.61 0.093 0.62 0.095
919 0.41 0.099 0.40 0.099
920 0.45 0.101 0.45 0.093
”21 0-26 0.073 0-27 0.072
Conditional Variance Parameters
(o m 0.00 €7.60 0.00
a
p 0.13 " 0.061 0.13 " 0.061
Skewness -0.33 -0.3 1
Kurtosis 5.71 5.64
02., 14.39 15.63
020’ 13.19 13.16
T 467 467

 

('0 Full period of 22 February, 2001 through 31 December, 2002.
(0 denotes 10% significance level

0'" denotes 5% significance level

(...) denotes 1% significance level

(51+k 'ft) = at + Z 6j£t_j 'i' b0 4' bIUStb + bZUStS
j=1,21

2 2 2
a, = a) + a£,_1 + ,BO',_1

gtzzta't with Zt is iid (0,1)

113

4.5 Conclusion

This chapter is concerned with the relation between the risk premium and central
bank intervention. Forward rates are calculated for the Turkish Lira-USS exchange
market and then the effect of central bank intervention on the risk premium is presented.
Using high quality daily intervention data from the Central Bank of Turkey as well as
implied forward rates, an MA(21)-GARCH(1,1) model is estimated. Both purchases and
sales of US dollars by the Central Bank of Turkey appear to have no effect on the size of
risk premium for TL/USD for the free ﬂoat period. Similar results are found for the
managed ﬂoat period. No empirical support is found for the theoretical model, with

intervention having a signiﬁcant effect on the risk premium.

114

CHAPTER 5

CONCLUSION

This dissertation provides an econometric analysis of the effects of and motivation for
central bank intervention in Turkey during the period 1993 - 2003. In the ﬁrst chapter, we
investigate the e ffect of intervention on both the level and volatility of spot returns in
Turkey. It is a common belief among policy-makers that central bank intervention is
effective. However, the empirical evidence is mixed, and indeed, tends to suggest
ineffectiveness overall. In this dissertation, we add to the existing empirical evidence, and
ﬁnd that central bank intervention is not effective in the Turkish foreign exchange market
during the period November 1St 1993 until December 31St 2003. We estimate a linear
GARCH( 1,1) model for different sub-periods including and excluding the crisis periods.
The results show that intervention does not affect the level of spot retm'ns both in the
managed and free ﬂoat regime except during the sub-period January 1St 2000 - February
21st 2001. While intervention has no effect on volatility in the managed ﬂoat regime, it
increases volatility in a free ﬂoat regime. With respect to the effect of the short-term
interest rate, an increase in overnight interest rates appreciates the US dollar in a

managed ﬂoat regime but has no effect in a free ﬂoat regime.

The second chapter investigates the motivation for central bank intervention. We
estimate the reaction function for both the managed ﬂoat and free ﬂoat regimes
separately. The central bank intervention reaction function is estimated using probit, tobit

models and then using the censored least absolute deviation method. The empirical

115

evidence gives different results for different regimes. First we estimated a probit model
for a managed ﬂoat and ﬁnd that the probability of selling the US dollar is increased by
the deviation from both the short and the medium term. The probability of a purchase, on
the other hand, is increased by the deviation from the long-term. Finally, the probability
of intervention is not determined by the conditional volatility for the managed ﬂoat. Our
tobit estimates give similar results. In contrast to a managed ﬂoat, both our tobit and
probit estimates suggest that the central bank is more responsive to volatility during a free
ﬂoat regime. While the sale of the US dollar is determined by volatility and the deviation
from the long-term, its purchase is determined by the deviation from the short and
medium term. When non-normality and heteroscedasticity in the tobit results are found,
all the reaction functions are estimated using the CLAD method so that we get
asymptotically normal and consistent estimators. For a free-ﬂoat regime, we ﬁnd that
neither the deviation from the 5, 50 and 125 days moving average values nor volatility
have any effect on the sales of the US dollar. In all of the estimations, the common
ﬁnding is that the lagged v alue o f i ntervention increases the n ext d ay intervention for

both a managed and free ﬂoat regime.

Chapter three examines the intervention reaction ﬁmction using the Threshold model
in the non-linear estimation of the reaction function. For a managed ﬂoat period, we fail
to accept both the two regime and three regime model. The linear models give that
deviation from the 5 days moving average and recent interventions matter but that
volatility does not. For a free ﬂoat period, under the assumption of homoscedastic errors,

we fail to accept linearity against the two regime model and both volatility and recent

116

interventions matter. These ﬁndings are consistent with the announced intention of the

CBRT not to affect the level of the exchange rate but rather the excess volatility.

Chapter four is concerned with the relation between the risk premium and central
bank intervention. Forward rates are calculated for the Turkish Lira-USS exchange
market and then the effect of central bank intervention on the risk premium is presented.
Using high quality daily intervention data from the Central Bank of Turkey as well as
implied forward rates, an MA(21)-GARCH(1,1) model is estimated. Both purchases and
sales of US dollars by the Central Bank of Turkey appear to have no effect on the size of
the risk premium for TL/USD for the free ﬂoat period. Similar results are found for the
managed ﬂoat period No empirical support is found for the theoretical model, with

intervention having a signiﬁcant effect on the risk premium.

117

REFERENCES

Almekinders, G. and Eij fﬁnger, S. (1996), “A friction model of daily Bundesbank and
Federal Reserve intervention”, Journal of Banking and Finance 20, 1365-1380.

Almekinders, 6.]. (1995), Foreign Exchange Intervention, Theory and Evidence,
Elgar Hants:UK.

Baillie, R.T., Lippens, RE. and McMahon, P.C.(1983), “Testing rational expectations
and efﬁciency in the foreign exchange market”, Econometrica 51, 553-563.

Baillie, RT (1989), “Econometric tests of rationality and market efﬁciency”,
Econometric Reviews 8, 151-186.

Baillie, R.T., Bollerslev, T. and Mikkelsen, H0. (1996), “Fractionally integrated
generalized autoregressive conditional heteroskedasticity”, Journal of Econometrics
74, 3-30.

Baillie, RT. and Osterberg, W.P. ( 1997a), “Central bank intervention and risk in the
forward market”, Journal of International Economics 43, 483-497.

Baillie, RT. and Osterberg, W.P. (1997b), “Why do central banks intervene?”, Journal of
International Money and Finance 16, 909-919.

Baillie, R.T., Humpage, CF. and Osterberg,W.P. (2000), “Intervention from an
information perspective”, Journal of International Financial Markets, Institutions
and Money 10, 407-421.

Baillie, R.T. and Osterberg, W.P. (2000), “Deviations from daily uncovered interest rate
parity and the role of intervention”, Journal of International Financial Markets,
Institutions and Money 10, 363-379.

Beine, M., Benassy-Quere, A. and Lecourt, C. (2002), “Central bank intervention and
foreign exchange rates; new evidence from F IGARCH estimations”, Journal of
International Money and Finance 21 , 115-144.

Bonser-Neal, C. and Tanner, G. (1996), “Central bank intervention and the volatility of

the foreign exchange rates: evidence from the options market”, Journal of
International Money and Finance 15, 853-878.

118

Bollerslev, T. (1986), “Generalized Autoregressive Conditional Heteroscedasticity”,
Journal of Econometrics, 31, 307-27.

Buchinsky, M. (1994), “Changes in the US wage structure 1963-1987: An application of
quantile regression.”Econometrica 62, 405-58.

Chang, Y. and Taylor, ST. (1998), “Intraday effects of foreign exchange intervention by
the Bank of Japan”, Journal of International Money and Finance 17, 191-210.

de Jong, R.M. and Herrera, A.M. (2004), “Dynamic censored regression on the open
market desk reaction function”, (unpublished manuscript).

de Jong, R.M. and Woutersen, T.M. (2004), “Dynamic time series binary choice”,
(unpublished manuscript).

Domac, I. and Mendoza, A. (2004) “Is there room for foreign exchange interventions
under an inﬂation targeting framework?”, World Bank Policy Research Working
Paper, No. 3288.

Dominquez, KM. (1990), “Market responses to coordinated central bank intervention”,
Carnegie-Rochester Conference on Public Policy 32, 121-64.

Dominquez, K. M. and Frankel, J.A. (1993), “Does foreign exchange intervention
matter? The Portfolio effect”, American Economic Review 83, 1356-1369.

Dominquez, KM. (1998), “Central bank intervention and exchange rate volatility”,
Journal of International Money and Finance 17, 161-190.

Dominquez, KM. (2003), “When do central bank interventions inﬂuence intra-daily and
longer-term exchange rate movements?”, NBER Working Paper Series No. 9875.

Edison, H. (1993), “The effectiveness of central bank intervention: a survey of the
literature after 1982”, Princeton University, Special Papers in International
Economics, No. 18.

Eij fﬁnger, S.C.W. and Gruijters, A.P.D. (1998), “On the short term objectiveness of daily
intervention by the Deutsche Bundesbanj and the federal reserve system in the US.
Dollar-Deutsche Mark exchange market”, in S.C.W. Eij fﬁnger, ed., Foreign
Exchange Intervention: Obiectivesﬁ and Effectiveness, Edward Elgar Publishing:
Cheltenham.

Eij fﬁnger, S. (1998), “The Ineffectiveness of central bank intervention”, in

S.C.W.Eij fﬁnger, ed., Foreigi Exchgge Intervention: Obiectives and Effectiveness”,
Edward Elgar Publishinngheltenham, 401-420.

119

 

Elliot, G. and Ito, T. (1995), “Heterogeneous expectations and tests of rationality in the
US Dollar/Y en forward foreign exchange rate market”, Working Paper, Department
of Economics, UC-San Diego.

Emir,O., Karasoy, A. and Kunter, K. (2000), “Monetary policy reaction function in
Turkey”, Research Department Working Paper, The Central bank of the Republic of

Turkey.

Evans, MD. and Lewis, K.K. (1995), “Do long term swings in the dollar affect estimates
of the risk premia?”, Review of Financial Studies 8, 709-742.

Fatum, R. and Hutchison, M.M. (2003), “Is sterilized foreign exchange intervention
effective after all? An event study approach”, The Economic Journal 113, 390-411.

Frankel, J .A and Froot, K.A. (1987), “Using survey data to test standard propositions
regarding exchange rate expectations”, American Economic Review 77, 133-153.

Frenkel, M., Pierdzioch, C. and Stadtrnann, G. (2003), “Modeling coordinated foreign
exchange market interventions: The case of the Japanese and US interventions in the
1990s”, Review of World Economics 139(4), 709-29.

Frenkel, M. and Stadtrnann, G. (2001), “Intervention reaction functions in the Dollar-
Deutshmark market”, Swiss Society for Financial Market Research 15, 328-343.

Galati, G. and Melick,W. (1999), “Perceived central bank intervention and market
expectations: an empirical study of the Yen-Dollar exchange rate”, BIS working
paper 77.

Geert, J .A. and Eij fﬁnger, S.C.W. (1994), “The ineffectiveness of central bank
intervention”, Center Discussion Paper Series, T ilburg University, No.94101,
December, 1-18.

Ghosh, A. (1992), “Is it signaling? Exchange intervention and the Dollar-Deutshemark
rate”, Journal of International Economics 32, 201 -220.

Gumus, I. (2001), “Effects of the interest rate defense on exchange rates during the 1994
crisis”, Research Department Working Paper, The Central Bank of the Republic of
Turkey 14, 1-17.

Grabbe, J .O. (1996), Intergtional Financial Markets, Third Edition, Prentice-HallzNew
York.

Hakkio, C. S. (1981), “Expectations and the forward exchange rate”, International
Economic Review 22, 663-678.

120

Hansen, LP. and Hodrick, R.J. (1980), “Forward exchange rates as optimal predictors of
future spot rates”, Journal of Political Economy 88, 829-853.

Hodrick, RT. (1989), “Risk, uncertainty and exchange rates”, Journal of Monetary
Economics 23, 433-459.

Hansen, BE. (1997), “Inference in TAR models”, Studies in Nonlinear Dynamics and
Econometrics 2, 1-14.

Hansen, BE. (1999), “Testing for linearity”, Journal of Economic Surveys 13, 551-576.

Hansen, BE. (2000), “Sample splitting and threshold estimation”, Econometrica 68, 575-
603.

Humpage, OF. (1988), “Intervention and the dollar’s decline”, Federal Reserve Bank of
Cleveland Econ. Rev. 3, 2-10.

Humpage, OR (1999), “US intervention: assessing the probability of success”, Journal
of Money and Credit Bank 31, 731-747.

Humpage, OF. (2003), “Government intervention in the foreign exchange market”,
Federal Reserve Bank of Cleveland Working Paper Series No. 03-15.

Hung, J. H. (1997), “Intervention strategies and exchange rate volatility : a noise trading
perspective”, Journal of International Money and Finance 16, 779-793.

Hutchison, M.M. (2003), “Intervention and exchange rate stabilization policy in
developing countries”, International Finance 6, 109-127.

Ito, T. (2002), “Is foreign exchange intervention effective? The Japanese experiences in
the 1990s”, NBER Working Paper Series, 8914.

Ito, T. and Yabu, T. (2004), “What promotes Japan to intervene in the Forex market? A
new approach to a reaction function”, NBER Working Paper Series, April 2004.

Jun, J. (2004), “Nonlinearity in central bank intervention in the DM/USD market” (Ph.D.
dissertation, Michigan State University).

Kaminsky, G. (1993), “Is there a Peso problem? Evidence from the Dollar/Pound
exchange rate”, American Economic Review 83, 450-472.

Khan, S., and Powell, J .L. ( 2001), “Two-step estimation of semiparametric censored
regression models”, Journal of Econometrics 103, 73-110.

121

Kim, S. and Sheen, J. (2002), “The determinants of foreign exchange intervention by
central banks: evidence from Australia”, Journal of International Money and
Finance 21, 619-649.

King, R.M. (2003), “Effective foreign exchange intervention: Matching strategies with
objectives”, International Finance 6, 249-271.

Leahy, M.P. (1995), “The proﬁtability of US intervention in the foreign exchange
markets”, Journal of International Money and Finance 14, 823-844.

Lewis, K.K. (1988), ”Testing the portfolio balance model: A multi-lateral approach”,
Journal of International Economics 24(1-2), 109-127.

Lewis, K.K. (1989), “The persistence of the Peso problem when policy is noisy”, Journal
of International Money and Finance 7, 1-11.

LeBaron, B.(1999), “Technical trading rule proﬁtability and foreign exchange
intervention”, Journal of International Economics 49, 125-143.

McKenzie, M. (2004), “An empirical examination of the relationship between central
bank intervention and exchange rate volatility : Some Australian evidence”,
Australian Economic Papers 43, 59-74.

Neeley, C. J. (1998), “Technical analysis and the proﬁtability of the US foreign exchange
intervention”, FED St. Louis Review July / August, 3-17.

Obstfeld, M.(1989), “The effectiveness of foreign exchange intervention”, NBER
working paper 2796.

Ozatay, F. (1994), “The Lessons from the 1994 crisis in Turkey: Public debt
(mis)management and conﬁdence crisis”, Yapi Kredi Economic Review 7, 21-37.

Ozatay, F. and Sak, G. (2002), “The ﬁnancial crisis in Turkey”, Central Bank of Turkey
and Ankara University, 1-29.

Popper, H. and Montgomery, J. (2001), “Information sharing and central bank
intervention in the foreign exchange market”, Journal of International Economics 55,
295-316.

Pagan, A. and Vella, F. (1989), “Diagnostic tests for models based on individual data: A
survey”, Journal of Applied Econometrics 4, $29-$59.

Powell, J .L. (1984), “Least absolute deviations estimation for the censored regression
model”, Journal of Econometrics 25, 303-325.

Riehl, H. and Rodriguez, R. (1977), Foreign Exchange Markets, McGraw HillzNew York.

122

 

Rogers, J. and Siklos, P. (2003), “Foreign exchange market intervention in two small
open economies: The Canadian and Australian experience”, Journal of International
Money and Finance 22, 393-416.

Saacke, P. (2002), “Technical analysis and the effectiveness of central bank
intervention”, Journal of International Money and Finance 21, 459-479.

Samo, L. and Taylor, M.P. (2001), “Ofﬁcial intervention in the foreign exchange market:
Is it effective, and, if so, how does it work?”, Centre for Economic Policy Research
Discussion Paper, No.2690.

Schwartz, J. A. (2000), “The rise and fall of foreign exchange market intervention”.
NBER Working Paper Series, No. 7751

Sweeney, R..J. (1997), “Do central banks lose on foreign exchange intervention?”,
Journal of Banking and Finance 21 , 1667-1684.

Vitale, P. (1999), “Sterilized central bank intervention in the foreign exchange market”,
Journal of International Economics 49, 245-267.

Wooldridge, J .M. (2000), “Introductory econometrics: A modern approach”,
South-Western College Publishinngincinnati.

Wooldridge, J .M. (2002), “Econometric analysis of cross section and panel data”,
CambridgezMIT Press.

Yeldan, E. (2001), “On the IMF-directed disinﬂation program in Turkey: A program for

stabilization and austerity or a recipe for impoverishments and ﬁnancial chaos?”,
Bilkent University:Turkey.

123

I!!! “ ‘ .... 111 ts

 

 

   

       

 

Mu’mrmu 5.7:.

11311111

1293 02